2008-11-10 21:21:49 +01:00
|
|
|
/* Copyright 2000-2008 MySQL AB, 2008 Sun Microsystems, Inc.
|
2000-08-28 15:43:58 +02:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
This program is free software; you can redistribute it and/or modify
|
|
|
|
it under the terms of the GNU General Public License as published by
|
2006-12-23 20:17:15 +01:00
|
|
|
the Free Software Foundation; version 2 of the License.
|
2000-08-28 15:43:58 +02:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
This program is distributed in the hope that it will be useful,
|
|
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
|
|
GNU General Public License for more details.
|
2000-08-28 15:43:58 +02:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
You should have received a copy of the GNU General Public License
|
|
|
|
along with this program; if not, write to the Free Software
|
|
|
|
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA */
|
|
|
|
|
|
|
|
|
2001-09-27 20:45:48 +02:00
|
|
|
/* Basic functions needed by many modules */
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
#include "mysql_priv.h"
|
2009-09-29 17:38:40 +02:00
|
|
|
#include "debug_sync.h"
|
2003-01-07 10:45:06 +01:00
|
|
|
#include "sql_select.h"
|
2004-09-07 14:29:46 +02:00
|
|
|
#include "sp_head.h"
|
2005-03-04 14:35:28 +01:00
|
|
|
#include "sp.h"
|
2004-09-07 14:29:46 +02:00
|
|
|
#include "sql_trigger.h"
|
2009-12-03 19:37:38 +01:00
|
|
|
#include "transaction.h"
|
Backport of revno 2630.28.10, 2630.28.31, 2630.28.26, 2630.33.1,
2630.39.1, 2630.28.29, 2630.34.3, 2630.34.2, 2630.34.1, 2630.29.29,
2630.29.28, 2630.31.1, 2630.28.13, 2630.28.10, 2617.23.14 and
some other minor revisions.
This patch implements:
WL#4264 "Backup: Stabilize Service Interface" -- all the
server prerequisites except si_objects.{h,cc} themselves (they can
be just copied over, when needed).
WL#4435: Support OUT-parameters in prepared statements.
(and all issues in the initial patches for these two
tasks, that were discovered in pushbuild and during testing).
Bug#39519: mysql_stmt_close() should flush all data
associated with the statement.
After execution of a prepared statement, send OUT parameters of the invoked
stored procedure, if any, to the client.
When using the binary protocol, send the parameters in an additional result
set over the wire. When using the text protocol, assign out parameters to
the user variables from the CALL(@var1, @var2, ...) specification.
The following refactoring has been made:
- Protocol::send_fields() was renamed to Protocol::send_result_set_metadata();
- A new Protocol::send_result_set_row() was introduced to incapsulate
common functionality for sending row data.
- Signature of Protocol::prepare_for_send() was changed: this operation
does not need a list of items, the number of items is fully sufficient.
The following backward incompatible changes have been made:
- CLIENT_MULTI_RESULTS is now enabled by default in the client;
- CLIENT_PS_MULTI_RESUTLS is now enabled by default in the client.
2009-10-21 22:02:06 +02:00
|
|
|
#include "sql_prepare.h"
|
2000-07-31 21:29:14 +02:00
|
|
|
#include <m_ctype.h>
|
|
|
|
#include <my_dir.h>
|
|
|
|
#include <hash.h>
|
2009-09-24 16:52:52 +02:00
|
|
|
#include "rpl_filter.h"
|
2006-05-30 15:07:49 +02:00
|
|
|
#ifdef __WIN__
|
2000-07-31 21:29:14 +02:00
|
|
|
#include <io.h>
|
|
|
|
#endif
|
|
|
|
|
2007-06-21 13:34:06 +02:00
|
|
|
#define FLAGSTR(S,F) ((S) & (F) ? #F " " : "")
|
|
|
|
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
/**
|
|
|
|
This internal handler is used to trap internally
|
|
|
|
errors that can occur when executing open table
|
|
|
|
during the prelocking phase.
|
|
|
|
*/
|
|
|
|
class Prelock_error_handler : public Internal_error_handler
|
|
|
|
{
|
|
|
|
public:
|
|
|
|
Prelock_error_handler()
|
|
|
|
: m_handled_errors(0), m_unhandled_errors(0)
|
|
|
|
{}
|
|
|
|
|
|
|
|
virtual ~Prelock_error_handler() {}
|
|
|
|
|
2009-09-10 11:18:29 +02:00
|
|
|
virtual bool handle_condition(THD *thd,
|
|
|
|
uint sql_errno,
|
|
|
|
const char* sqlstate,
|
|
|
|
MYSQL_ERROR::enum_warning_level level,
|
|
|
|
const char* msg,
|
|
|
|
MYSQL_ERROR ** cond_hdl);
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
|
|
|
|
bool safely_trapped_errors();
|
|
|
|
|
|
|
|
private:
|
|
|
|
int m_handled_errors;
|
|
|
|
int m_unhandled_errors;
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
|
|
bool
|
2009-09-10 11:18:29 +02:00
|
|
|
Prelock_error_handler::handle_condition(THD *,
|
|
|
|
uint sql_errno,
|
|
|
|
const char*,
|
|
|
|
MYSQL_ERROR::enum_warning_level,
|
|
|
|
const char*,
|
|
|
|
MYSQL_ERROR ** cond_hdl)
|
|
|
|
{
|
|
|
|
*cond_hdl= NULL;
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
if (sql_errno == ER_NO_SUCH_TABLE)
|
|
|
|
{
|
|
|
|
m_handled_errors++;
|
2007-07-20 17:46:13 +02:00
|
|
|
return TRUE;
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
m_unhandled_errors++;
|
2007-07-20 17:46:13 +02:00
|
|
|
return FALSE;
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
bool Prelock_error_handler::safely_trapped_errors()
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
If m_unhandled_errors != 0, something else, unanticipated, happened,
|
|
|
|
so the error is not trapped but returned to the caller.
|
|
|
|
Multiple ER_NO_SUCH_TABLE can be raised in case of views.
|
|
|
|
*/
|
|
|
|
return ((m_handled_errors > 0) && (m_unhandled_errors == 0));
|
|
|
|
}
|
|
|
|
|
2007-08-15 17:08:44 +02:00
|
|
|
/**
|
|
|
|
@defgroup Data_Dictionary Data Dictionary
|
|
|
|
@{
|
|
|
|
*/
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
/**
|
|
|
|
Total number of TABLE instances for tables in the table definition cache
|
|
|
|
(both in use by threads and not in use). This value is accessible to user
|
|
|
|
as "Open_tables" status variable.
|
|
|
|
*/
|
|
|
|
uint table_cache_count= 0;
|
|
|
|
/**
|
|
|
|
List that contains all TABLE instances for tables in the table definition
|
|
|
|
cache that are not in use by any thread. Recently used TABLE instances are
|
|
|
|
appended to the end of the list. Thus the beginning of the list contains
|
|
|
|
tables which have been least recently used.
|
|
|
|
*/
|
|
|
|
TABLE *unused_tables;
|
|
|
|
HASH table_def_cache;
|
2005-11-23 21:45:02 +01:00
|
|
|
static TABLE_SHARE *oldest_unused_share, end_of_unused_share;
|
|
|
|
static bool table_def_inited= 0;
|
2009-12-08 09:26:49 +01:00
|
|
|
static bool table_def_shutdown_in_progress= 0;
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
static bool check_and_update_table_version(THD *thd, TABLE_LIST *tables,
|
|
|
|
TABLE_SHARE *table_share);
|
|
|
|
static bool open_table_entry_fini(THD *thd, TABLE_SHARE *share, TABLE *entry);
|
|
|
|
static bool auto_repair_table(THD *thd, TABLE_LIST *table_list);
|
2000-07-31 21:29:14 +02:00
|
|
|
static void free_cache_entry(TABLE *entry);
|
2005-11-23 21:45:02 +01:00
|
|
|
static bool open_new_frm(THD *thd, TABLE_SHARE *share, const char *alias,
|
2005-11-01 14:54:30 +01:00
|
|
|
uint db_stat, uint prgflag,
|
|
|
|
uint ha_open_flags, TABLE *outparam,
|
|
|
|
TABLE_LIST *table_desc, MEM_ROOT *mem_root);
|
2009-12-08 10:57:07 +01:00
|
|
|
static bool tdc_wait_for_old_versions(THD *thd,
|
|
|
|
MDL_request_list *mdl_requests);
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
static bool
|
2009-10-01 01:19:36 +02:00
|
|
|
has_write_table_with_auto_increment(TABLE_LIST *tables);
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-12-09 10:44:01 +01:00
|
|
|
TABLE *find_table_for_mdl_upgrade(TABLE *list, const char *db,
|
|
|
|
const char *table_name);
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
uint cached_open_tables(void)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
return table_cache_count;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
#ifdef EXTRA_DEBUG
|
|
|
|
static void check_unused(void)
|
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
uint count= 0, open_files= 0, idx= 0;
|
2009-11-30 16:55:03 +01:00
|
|
|
TABLE *cur_link, *start_link, *entry;
|
|
|
|
TABLE_SHARE *share;
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
if ((start_link=cur_link=unused_tables))
|
|
|
|
{
|
|
|
|
do
|
|
|
|
{
|
|
|
|
if (cur_link != cur_link->next->prev || cur_link != cur_link->prev->next)
|
|
|
|
{
|
|
|
|
DBUG_PRINT("error",("Unused_links aren't linked properly")); /* purecov: inspected */
|
|
|
|
return; /* purecov: inspected */
|
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
} while (count++ < table_cache_count &&
|
2000-07-31 21:29:14 +02:00
|
|
|
(cur_link=cur_link->next) != start_link);
|
|
|
|
if (cur_link != start_link)
|
|
|
|
{
|
|
|
|
DBUG_PRINT("error",("Unused_links aren't connected")); /* purecov: inspected */
|
|
|
|
}
|
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
for (idx=0 ; idx < table_def_cache.records ; idx++)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
share= (TABLE_SHARE*) my_hash_element(&table_def_cache, idx);
|
|
|
|
|
|
|
|
I_P_List_iterator<TABLE, TABLE_share> it(share->free_tables);
|
|
|
|
while ((entry= it++))
|
|
|
|
{
|
2009-12-03 00:09:22 +01:00
|
|
|
/* We must not have TABLEs in the free list that have their file closed. */
|
|
|
|
DBUG_ASSERT(entry->db_stat && entry->file);
|
|
|
|
/* Merge children should be detached from a merge parent */
|
|
|
|
DBUG_ASSERT(! entry->file->extra(HA_EXTRA_IS_ATTACHED_CHILDREN));
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
if (entry->in_use)
|
|
|
|
{
|
|
|
|
DBUG_PRINT("error",("Used table is in share's list of unused tables")); /* purecov: inspected */
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
count--;
|
2005-11-23 21:45:02 +01:00
|
|
|
open_files++;
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
|
|
|
it.init(share->used_tables);
|
|
|
|
while ((entry= it++))
|
|
|
|
{
|
|
|
|
if (!entry->in_use)
|
|
|
|
{
|
|
|
|
DBUG_PRINT("error",("Unused table is in share's list of used tables")); /* purecov: inspected */
|
|
|
|
}
|
|
|
|
open_files++;
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
if (count != 0)
|
|
|
|
{
|
2000-08-16 11:17:40 +02:00
|
|
|
DBUG_PRINT("error",("Unused_links doesn't match open_cache: diff: %d", /* purecov: inspected */
|
2000-07-31 21:29:14 +02:00
|
|
|
count)); /* purecov: inspected */
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
#define check_unused()
|
|
|
|
#endif
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
|
|
|
|
/*
|
|
|
|
Create a table cache key
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
create_table_def_key()
|
|
|
|
thd Thread handler
|
|
|
|
key Create key here (must be of size MAX_DBKEY_LENGTH)
|
|
|
|
table_list Table definition
|
|
|
|
tmp_table Set if table is a tmp table
|
|
|
|
|
|
|
|
IMPLEMENTATION
|
|
|
|
The table cache_key is created from:
|
|
|
|
db_name + \0
|
|
|
|
table_name + \0
|
|
|
|
|
|
|
|
if the table is a tmp table, we add the following to make each tmp table
|
|
|
|
unique on the slave:
|
|
|
|
|
|
|
|
4 bytes for master thread id
|
|
|
|
4 bytes pseudo thread id
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
Length of key
|
|
|
|
*/
|
|
|
|
|
2005-11-29 09:07:21 +01:00
|
|
|
uint create_table_def_key(THD *thd, char *key, TABLE_LIST *table_list,
|
2005-11-23 21:45:02 +01:00
|
|
|
bool tmp_table)
|
|
|
|
{
|
|
|
|
uint key_length= (uint) (strmov(strmov(key, table_list->db)+1,
|
|
|
|
table_list->table_name)-key)+1;
|
|
|
|
if (tmp_table)
|
|
|
|
{
|
|
|
|
int4store(key + key_length, thd->server_id);
|
|
|
|
int4store(key + key_length + 4, thd->variables.pseudo_thread_id);
|
|
|
|
key_length+= TMP_TABLE_KEY_EXTRA;
|
|
|
|
}
|
|
|
|
return key_length;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/*****************************************************************************
|
|
|
|
Functions to handle table definition cach (TABLE_SHARE)
|
|
|
|
*****************************************************************************/
|
|
|
|
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
extern "C" uchar *table_def_key(const uchar *record, size_t *length,
|
2005-11-23 21:45:02 +01:00
|
|
|
my_bool not_used __attribute__((unused)))
|
|
|
|
{
|
|
|
|
TABLE_SHARE *entry=(TABLE_SHARE*) record;
|
|
|
|
*length= entry->table_cache_key.length;
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
return (uchar*) entry->table_cache_key.str;
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static void table_def_free_entry(TABLE_SHARE *share)
|
|
|
|
{
|
|
|
|
DBUG_ENTER("table_def_free_entry");
|
2009-11-30 20:38:25 +01:00
|
|
|
safe_mutex_assert_owner(&LOCK_open);
|
2005-11-23 21:45:02 +01:00
|
|
|
if (share->prev)
|
|
|
|
{
|
|
|
|
/* remove from old_unused_share list */
|
|
|
|
*share->prev= share->next;
|
|
|
|
share->next->prev= share->prev;
|
|
|
|
}
|
|
|
|
free_table_share(share);
|
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
bool table_def_init(void)
|
|
|
|
{
|
|
|
|
table_def_inited= 1;
|
|
|
|
oldest_unused_share= &end_of_unused_share;
|
|
|
|
end_of_unused_share.prev= &oldest_unused_share;
|
|
|
|
|
2009-10-14 18:37:38 +02:00
|
|
|
return my_hash_init(&table_def_cache, &my_charset_bin, table_def_size,
|
|
|
|
0, 0, table_def_key,
|
|
|
|
(my_hash_free_key) table_def_free_entry, 0) != 0;
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-08 09:26:49 +01:00
|
|
|
/**
|
|
|
|
Notify table definition cache that process of shutting down server
|
|
|
|
has started so it has to keep number of TABLE and TABLE_SHARE objects
|
|
|
|
minimal in order to reduce number of references to pluggable engines.
|
|
|
|
*/
|
|
|
|
|
|
|
|
void table_def_start_shutdown(void)
|
|
|
|
{
|
|
|
|
if (table_def_inited)
|
|
|
|
{
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
/* Free all cached but unused TABLEs and TABLE_SHAREs first. */
|
|
|
|
close_cached_tables(NULL, NULL, TRUE, FALSE);
|
|
|
|
/*
|
|
|
|
Ensure that TABLE and TABLE_SHARE objects which are created for
|
|
|
|
tables that are open during process of plugins' shutdown are
|
|
|
|
immediately released. This keeps number of references to engine
|
|
|
|
plugins minimal and allows shutdown to proceed smoothly.
|
|
|
|
*/
|
|
|
|
table_def_shutdown_in_progress= TRUE;
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
void table_def_free(void)
|
|
|
|
{
|
|
|
|
DBUG_ENTER("table_def_free");
|
|
|
|
if (table_def_inited)
|
|
|
|
{
|
|
|
|
table_def_inited= 0;
|
2009-11-30 16:55:03 +01:00
|
|
|
/* Free table definitions. */
|
2009-10-14 18:37:38 +02:00
|
|
|
my_hash_free(&table_def_cache);
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
uint cached_table_definitions(void)
|
|
|
|
{
|
|
|
|
return table_def_cache.records;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/*
|
|
|
|
Auxiliary routines for manipulating with per-share used/unused and
|
|
|
|
global unused lists of TABLE objects and table_cache_count counter.
|
|
|
|
Responsible for preserving invariants between those lists, counter
|
|
|
|
and TABLE::in_use member.
|
|
|
|
In fact those routines implement sort of implicit table cache as
|
|
|
|
part of table definition cache.
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Add newly created TABLE object for table share which is going
|
|
|
|
to be used right away.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void table_def_add_used_table(THD *thd, TABLE *table)
|
|
|
|
{
|
|
|
|
DBUG_ASSERT(table->in_use == thd);
|
|
|
|
table->s->used_tables.push_front(table);
|
|
|
|
table_cache_count++;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Prepare used or unused TABLE instance for destruction by removing
|
|
|
|
it from share's and global list.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void table_def_remove_table(TABLE *table)
|
|
|
|
{
|
|
|
|
if (table->in_use)
|
|
|
|
{
|
|
|
|
/* Remove from per-share chain of used TABLE objects. */
|
|
|
|
table->s->used_tables.remove(table);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/* Remove from per-share chain of unused TABLE objects. */
|
|
|
|
table->s->free_tables.remove(table);
|
|
|
|
|
|
|
|
/* And global unused chain. */
|
|
|
|
table->next->prev=table->prev;
|
|
|
|
table->prev->next=table->next;
|
|
|
|
if (table == unused_tables)
|
|
|
|
{
|
|
|
|
unused_tables=unused_tables->next;
|
|
|
|
if (table == unused_tables)
|
|
|
|
unused_tables=0;
|
|
|
|
}
|
|
|
|
check_unused();
|
|
|
|
}
|
|
|
|
table_cache_count--;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Mark already existing TABLE instance as used.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void table_def_use_table(THD *thd, TABLE *table)
|
|
|
|
{
|
|
|
|
DBUG_ASSERT(!table->in_use);
|
|
|
|
|
|
|
|
/* Unlink table from list of unused tables for this share. */
|
|
|
|
table->s->free_tables.remove(table);
|
|
|
|
/* Unlink able from global unused tables list. */
|
|
|
|
if (table == unused_tables)
|
|
|
|
{ // First unused
|
|
|
|
unused_tables=unused_tables->next; // Remove from link
|
|
|
|
if (table == unused_tables)
|
|
|
|
unused_tables=0;
|
|
|
|
}
|
|
|
|
table->prev->next=table->next; /* Remove from unused list */
|
|
|
|
table->next->prev=table->prev;
|
|
|
|
check_unused();
|
|
|
|
/* Add table to list of used tables for this share. */
|
|
|
|
table->s->used_tables.push_front(table);
|
|
|
|
table->in_use= thd;
|
2009-12-03 00:09:22 +01:00
|
|
|
/* The ex-unused table must be fully functional. */
|
|
|
|
DBUG_ASSERT(table->db_stat && table->file);
|
|
|
|
/* The children must be detached from the table. */
|
|
|
|
DBUG_ASSERT(! table->file->extra(HA_EXTRA_IS_ATTACHED_CHILDREN));
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Mark already existing used TABLE instance as unused.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void table_def_unuse_table(TABLE *table)
|
|
|
|
{
|
|
|
|
DBUG_ASSERT(table->in_use);
|
|
|
|
|
|
|
|
table->in_use= 0;
|
|
|
|
/* Remove table from the list of tables used in this share. */
|
|
|
|
table->s->used_tables.remove(table);
|
|
|
|
/* Add table to the list of unused TABLE objects for this share. */
|
|
|
|
table->s->free_tables.push_front(table);
|
|
|
|
/* Also link it last in the global list of unused TABLE objects. */
|
|
|
|
if (unused_tables)
|
|
|
|
{
|
|
|
|
table->next=unused_tables;
|
|
|
|
table->prev=unused_tables->prev;
|
|
|
|
unused_tables->prev=table;
|
|
|
|
table->prev->next=table;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
unused_tables=table->next=table->prev=table;
|
|
|
|
check_unused();
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
/*
|
|
|
|
Get TABLE_SHARE for a table.
|
|
|
|
|
|
|
|
get_table_share()
|
2006-08-02 17:57:06 +02:00
|
|
|
thd Thread handle
|
2005-11-23 21:45:02 +01:00
|
|
|
table_list Table that should be opened
|
|
|
|
key Table cache key
|
|
|
|
key_length Length of key
|
|
|
|
db_flags Flags to open_table_def():
|
|
|
|
OPEN_VIEW
|
|
|
|
error out: Error code from open_table_def()
|
|
|
|
|
|
|
|
IMPLEMENTATION
|
|
|
|
Get a table definition from the table definition cache.
|
|
|
|
If it doesn't exist, create a new from the table definition file.
|
|
|
|
|
|
|
|
NOTES
|
|
|
|
We must have wrlock on LOCK_open when we come here
|
|
|
|
(To be changed later)
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
0 Error
|
|
|
|
# Share for table
|
|
|
|
*/
|
|
|
|
|
2005-11-29 09:07:21 +01:00
|
|
|
TABLE_SHARE *get_table_share(THD *thd, TABLE_LIST *table_list, char *key,
|
2005-11-23 21:45:02 +01:00
|
|
|
uint key_length, uint db_flags, int *error)
|
|
|
|
{
|
|
|
|
TABLE_SHARE *share;
|
|
|
|
DBUG_ENTER("get_table_share");
|
|
|
|
|
|
|
|
*error= 0;
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/*
|
|
|
|
To be able perform any operation on table we should own
|
|
|
|
some kind of metadata lock on it.
|
|
|
|
*/
|
2009-12-10 09:21:38 +01:00
|
|
|
DBUG_ASSERT(thd->mdl_context.is_lock_owner(MDL_key::TABLE, table_list->db,
|
2009-12-04 00:52:05 +01:00
|
|
|
table_list->table_name));
|
2009-11-30 16:55:03 +01:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
/* Read table definition from cache */
|
2009-10-14 18:37:38 +02:00
|
|
|
if ((share= (TABLE_SHARE*) my_hash_search(&table_def_cache,(uchar*) key,
|
|
|
|
key_length)))
|
2005-11-23 21:45:02 +01:00
|
|
|
goto found;
|
|
|
|
|
|
|
|
if (!(share= alloc_table_share(table_list, key, key_length)))
|
|
|
|
{
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
|
2006-02-23 13:34:03 +01:00
|
|
|
/*
|
|
|
|
We assign a new table id under the protection of the LOCK_open and
|
|
|
|
the share's own mutex. We do this insted of creating a new mutex
|
|
|
|
and using it for the sole purpose of serializing accesses to a
|
|
|
|
static variable, we assign the table id here. We assign it to the
|
|
|
|
share before inserting it into the table_def_cache to be really
|
|
|
|
sure that it cannot be read from the cache without having a table
|
|
|
|
id assigned.
|
|
|
|
|
|
|
|
CAVEAT. This means that the table cannot be used for
|
|
|
|
binlogging/replication purposes, unless get_table_share() has been
|
|
|
|
called directly or indirectly.
|
|
|
|
*/
|
|
|
|
assign_new_table_id(share);
|
|
|
|
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
if (my_hash_insert(&table_def_cache, (uchar*) share))
|
2005-11-23 21:45:02 +01:00
|
|
|
{
|
|
|
|
free_table_share(share);
|
|
|
|
DBUG_RETURN(0); // return error
|
|
|
|
}
|
|
|
|
if (open_table_def(thd, share, db_flags))
|
|
|
|
{
|
|
|
|
*error= share->error;
|
2009-10-14 18:37:38 +02:00
|
|
|
(void) my_hash_delete(&table_def_cache, (uchar*) share);
|
2005-11-23 21:45:02 +01:00
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
share->ref_count++; // Mark in use
|
|
|
|
DBUG_PRINT("exit", ("share: 0x%lx ref_count: %u",
|
|
|
|
(ulong) share, share->ref_count));
|
|
|
|
DBUG_RETURN(share);
|
|
|
|
|
|
|
|
found:
|
|
|
|
/*
|
|
|
|
We found an existing table definition. Return it if we didn't get
|
|
|
|
an error when reading the table definition from file.
|
|
|
|
*/
|
|
|
|
if (share->error)
|
|
|
|
{
|
|
|
|
/* Table definition contained an error */
|
|
|
|
open_table_error(share, share->error, share->open_errno, share->errarg);
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
if (share->is_view && !(db_flags & OPEN_VIEW))
|
|
|
|
{
|
|
|
|
open_table_error(share, 1, ENOENT, 0);
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!share->ref_count++ && share->prev)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
Share was not used before and it was in the old_unused_share list
|
|
|
|
Unlink share from this list
|
|
|
|
*/
|
|
|
|
DBUG_PRINT("info", ("Unlinking from not used list"));
|
|
|
|
*share->prev= share->next;
|
|
|
|
share->next->prev= share->prev;
|
|
|
|
share->next= 0;
|
|
|
|
share->prev= 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Free cache if too big */
|
|
|
|
while (table_def_cache.records > table_def_size &&
|
|
|
|
oldest_unused_share->next)
|
2009-11-24 14:54:59 +01:00
|
|
|
my_hash_delete(&table_def_cache, (uchar*) oldest_unused_share);
|
2005-11-23 21:45:02 +01:00
|
|
|
|
|
|
|
DBUG_PRINT("exit", ("share: 0x%lx ref_count: %u",
|
|
|
|
(ulong) share, share->ref_count));
|
|
|
|
DBUG_RETURN(share);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Get a table share. If it didn't exist, try creating it from engine
|
|
|
|
|
|
|
|
For arguments and return values, see get_table_from_share()
|
|
|
|
*/
|
|
|
|
|
|
|
|
static TABLE_SHARE
|
|
|
|
*get_table_share_with_create(THD *thd, TABLE_LIST *table_list,
|
2005-11-29 09:07:21 +01:00
|
|
|
char *key, uint key_length,
|
2005-11-23 21:45:02 +01:00
|
|
|
uint db_flags, int *error)
|
|
|
|
{
|
|
|
|
TABLE_SHARE *share;
|
|
|
|
int tmp;
|
|
|
|
DBUG_ENTER("get_table_share_with_create");
|
|
|
|
|
2007-12-12 16:21:01 +01:00
|
|
|
share= get_table_share(thd, table_list, key, key_length, db_flags, error);
|
|
|
|
/*
|
|
|
|
If share is not NULL, we found an existing share.
|
|
|
|
|
|
|
|
If share is NULL, and there is no error, we're inside
|
|
|
|
pre-locking, which silences 'ER_NO_SUCH_TABLE' errors
|
|
|
|
with the intention to silently drop non-existing tables
|
|
|
|
from the pre-locking list. In this case we still need to try
|
|
|
|
auto-discover before returning a NULL share.
|
|
|
|
|
|
|
|
If share is NULL and the error is ER_NO_SUCH_TABLE, this is
|
|
|
|
the same as above, only that the error was not silenced by
|
|
|
|
pre-locking. Once again, we need to try to auto-discover
|
|
|
|
the share.
|
|
|
|
|
|
|
|
Finally, if share is still NULL, it's a real error and we need
|
|
|
|
to abort.
|
|
|
|
|
|
|
|
@todo Rework alternative ways to deal with ER_NO_SUCH TABLE.
|
|
|
|
*/
|
2009-09-10 11:18:29 +02:00
|
|
|
if (share || (thd->is_error() && thd->stmt_da->sql_errno() != ER_NO_SUCH_TABLE))
|
2007-12-12 16:21:01 +01:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
DBUG_RETURN(share);
|
|
|
|
|
|
|
|
/* Table didn't exist. Check if some engine can provide it */
|
2009-06-17 16:56:44 +02:00
|
|
|
tmp= ha_create_table_from_engine(thd, table_list->db,
|
|
|
|
table_list->table_name);
|
|
|
|
if (tmp < 0)
|
2005-11-23 21:45:02 +01:00
|
|
|
{
|
|
|
|
/*
|
|
|
|
No such table in any engine.
|
2007-12-12 16:21:01 +01:00
|
|
|
Hide "Table doesn't exist" errors if the table belongs to a view.
|
|
|
|
The check for thd->is_error() is necessary to not push an
|
|
|
|
unwanted error in case of pre-locking, which silences
|
|
|
|
"no such table" errors.
|
|
|
|
@todo Rework the alternative ways to deal with ER_NO_SUCH TABLE.
|
2005-11-23 21:45:02 +01:00
|
|
|
*/
|
2008-10-09 10:55:16 +02:00
|
|
|
if (thd->is_error())
|
2005-11-23 21:45:02 +01:00
|
|
|
{
|
2008-10-09 10:55:16 +02:00
|
|
|
if (table_list->parent_l)
|
|
|
|
{
|
|
|
|
thd->clear_error();
|
|
|
|
my_error(ER_WRONG_MRG_TABLE, MYF(0));
|
|
|
|
}
|
|
|
|
else if (table_list->belong_to_view)
|
|
|
|
{
|
|
|
|
TABLE_LIST *view= table_list->belong_to_view;
|
|
|
|
thd->clear_error();
|
|
|
|
my_error(ER_VIEW_INVALID, MYF(0),
|
|
|
|
view->view_db.str, view->view_name.str);
|
|
|
|
}
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
if (tmp)
|
|
|
|
{
|
|
|
|
/* Give right error message */
|
|
|
|
thd->clear_error();
|
|
|
|
DBUG_PRINT("error", ("Discovery of %s/%s failed", table_list->db,
|
|
|
|
table_list->table_name));
|
|
|
|
my_printf_error(ER_UNKNOWN_ERROR,
|
|
|
|
"Failed to open '%-.64s', error while "
|
|
|
|
"unpacking from engine",
|
|
|
|
MYF(0), table_list->table_name);
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
/* Table existed in engine. Let's open it */
|
2009-09-10 11:18:29 +02:00
|
|
|
thd->warning_info->clear_warning_info(thd->query_id);
|
2005-11-23 21:45:02 +01:00
|
|
|
thd->clear_error(); // Clear error message
|
|
|
|
DBUG_RETURN(get_table_share(thd, table_list, key, key_length,
|
|
|
|
db_flags, error));
|
|
|
|
}
|
|
|
|
|
2009-11-30 20:11:32 +01:00
|
|
|
/**
|
|
|
|
Mark that we are not using table share anymore.
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-11-30 20:11:32 +01:00
|
|
|
@param share Table share
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-11-30 20:11:32 +01:00
|
|
|
If the share has no open tables and (we have done a refresh or
|
|
|
|
if we have already too many open table shares) then delete the
|
|
|
|
definition.
|
2005-11-23 21:45:02 +01:00
|
|
|
*/
|
|
|
|
|
2009-11-30 20:11:32 +01:00
|
|
|
void release_table_share(TABLE_SHARE *share)
|
2005-11-23 21:45:02 +01:00
|
|
|
{
|
|
|
|
bool to_be_deleted= 0;
|
|
|
|
DBUG_ENTER("release_table_share");
|
|
|
|
DBUG_PRINT("enter",
|
|
|
|
("share: 0x%lx table: %s.%s ref_count: %u version: %lu",
|
|
|
|
(ulong) share, share->db.str, share->table_name.str,
|
|
|
|
share->ref_count, share->version));
|
|
|
|
|
|
|
|
safe_mutex_assert_owner(&LOCK_open);
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
DBUG_ASSERT(share->ref_count);
|
2005-11-23 21:45:02 +01:00
|
|
|
if (!--share->ref_count)
|
|
|
|
{
|
2009-12-08 09:26:49 +01:00
|
|
|
if (share->version != refresh_version ||
|
|
|
|
table_def_shutdown_in_progress)
|
2005-11-23 21:45:02 +01:00
|
|
|
to_be_deleted=1;
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/* Link share last in used_table_share list */
|
|
|
|
DBUG_PRINT("info",("moving share to unused list"));
|
|
|
|
|
|
|
|
DBUG_ASSERT(share->next == 0);
|
|
|
|
share->prev= end_of_unused_share.prev;
|
|
|
|
*end_of_unused_share.prev= share;
|
|
|
|
end_of_unused_share.prev= &share->next;
|
|
|
|
share->next= &end_of_unused_share;
|
|
|
|
|
|
|
|
to_be_deleted= (table_def_cache.records > table_def_size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (to_be_deleted)
|
|
|
|
{
|
|
|
|
DBUG_PRINT("info", ("Deleting share"));
|
2009-10-14 18:37:38 +02:00
|
|
|
my_hash_delete(&table_def_cache, (uchar*) share);
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Check if table definition exits in cache
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
get_cached_table_share()
|
|
|
|
db Database name
|
|
|
|
table_name Table name
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
0 Not cached
|
|
|
|
# TABLE_SHARE for table
|
|
|
|
*/
|
|
|
|
|
|
|
|
TABLE_SHARE *get_cached_table_share(const char *db, const char *table_name)
|
|
|
|
{
|
2005-11-29 09:07:21 +01:00
|
|
|
char key[NAME_LEN*2+2];
|
2005-11-23 21:45:02 +01:00
|
|
|
TABLE_LIST table_list;
|
|
|
|
uint key_length;
|
|
|
|
safe_mutex_assert_owner(&LOCK_open);
|
|
|
|
|
|
|
|
table_list.db= (char*) db;
|
|
|
|
table_list.table_name= (char*) table_name;
|
|
|
|
key_length= create_table_def_key((THD*) 0, key, &table_list, 0);
|
2009-10-14 18:37:38 +02:00
|
|
|
return (TABLE_SHARE*) my_hash_search(&table_def_cache,
|
|
|
|
(uchar*) key, key_length);
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/**
|
|
|
|
@brief Mark table share as having one more user (increase its reference
|
|
|
|
count).
|
|
|
|
|
|
|
|
@param share Table share for which reference count should be increased.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void reference_table_share(TABLE_SHARE *share)
|
|
|
|
{
|
|
|
|
DBUG_ENTER("reference_table_share");
|
|
|
|
DBUG_ASSERT(share->ref_count);
|
2009-11-30 20:38:25 +01:00
|
|
|
safe_mutex_assert_owner(&LOCK_open);
|
2009-11-30 16:55:03 +01:00
|
|
|
share->ref_count++;
|
|
|
|
DBUG_PRINT("exit", ("share: 0x%lx ref_count: %u",
|
|
|
|
(ulong) share, share->ref_count));
|
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2002-08-08 15:41:04 +02:00
|
|
|
/*
|
|
|
|
Create a list for all open tables matching SQL expression
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
list_open_tables()
|
|
|
|
thd Thread THD
|
|
|
|
wild SQL like expression
|
|
|
|
|
|
|
|
NOTES
|
|
|
|
One gets only a list of tables for which one has any kind of privilege.
|
|
|
|
db and table names are allocated in result struct, so one doesn't need
|
|
|
|
a lock on LOCK_open when traversing the return list.
|
|
|
|
|
|
|
|
RETURN VALUES
|
|
|
|
NULL Error (Probably OOM)
|
|
|
|
# Pointer to list of names of open tables.
|
|
|
|
*/
|
|
|
|
|
2005-08-16 11:18:35 +02:00
|
|
|
OPEN_TABLE_LIST *list_open_tables(THD *thd, const char *db, const char *wild)
|
2000-12-18 22:22:20 +01:00
|
|
|
{
|
|
|
|
int result = 0;
|
2001-07-11 09:36:22 +02:00
|
|
|
OPEN_TABLE_LIST **start_list, *open_list;
|
2000-12-18 22:22:20 +01:00
|
|
|
TABLE_LIST table_list;
|
|
|
|
DBUG_ENTER("list_open_tables");
|
2001-07-11 09:36:22 +02:00
|
|
|
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_lock(&LOCK_open);
|
2000-12-18 22:22:20 +01:00
|
|
|
bzero((char*) &table_list,sizeof(table_list));
|
2001-07-11 09:36:22 +02:00
|
|
|
start_list= &open_list;
|
|
|
|
open_list=0;
|
2000-12-18 22:22:20 +01:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
for (uint idx=0 ; result == 0 && idx < table_def_cache.records; idx++)
|
2000-12-18 22:22:20 +01:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
TABLE_SHARE *share= (TABLE_SHARE *)my_hash_element(&table_def_cache, idx);
|
2001-07-11 09:36:22 +02:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
if (db && my_strcasecmp(system_charset_info, db, share->db.str))
|
2005-08-20 10:00:00 +02:00
|
|
|
continue;
|
2005-11-23 21:45:02 +01:00
|
|
|
if (wild && wild_compare(share->table_name.str, wild, 0))
|
2005-08-16 11:18:35 +02:00
|
|
|
continue;
|
2000-12-18 22:22:20 +01:00
|
|
|
|
2001-07-11 09:36:22 +02:00
|
|
|
/* Check if user has SELECT privilege for any column in the table */
|
2005-11-23 21:45:02 +01:00
|
|
|
table_list.db= share->db.str;
|
|
|
|
table_list.table_name= share->table_name.str;
|
2001-07-11 09:36:22 +02:00
|
|
|
table_list.grant.privilege=0;
|
2004-06-25 15:52:01 +02:00
|
|
|
|
2009-10-19 14:58:13 +02:00
|
|
|
if (check_table_access(thd,SELECT_ACL,&table_list, TRUE, 1, TRUE))
|
2001-07-11 09:36:22 +02:00
|
|
|
continue;
|
2009-11-30 16:55:03 +01:00
|
|
|
|
2001-07-11 09:36:22 +02:00
|
|
|
if (!(*start_list = (OPEN_TABLE_LIST *)
|
2005-11-23 21:45:02 +01:00
|
|
|
sql_alloc(sizeof(**start_list)+share->table_cache_key.length)))
|
2000-12-18 22:22:20 +01:00
|
|
|
{
|
2001-07-11 09:36:22 +02:00
|
|
|
open_list=0; // Out of memory
|
2001-04-11 13:04:03 +02:00
|
|
|
break;
|
2000-12-18 22:22:20 +01:00
|
|
|
}
|
2001-09-27 20:45:48 +02:00
|
|
|
strmov((*start_list)->table=
|
|
|
|
strmov(((*start_list)->db= (char*) ((*start_list)+1)),
|
2005-11-23 21:45:02 +01:00
|
|
|
share->db.str)+1,
|
|
|
|
share->table_name.str);
|
2009-11-30 16:55:03 +01:00
|
|
|
(*start_list)->in_use= 0;
|
|
|
|
I_P_List_iterator<TABLE, TABLE_share> it(share->used_tables);
|
|
|
|
while (it++)
|
|
|
|
++(*start_list)->in_use;
|
|
|
|
(*start_list)->locked= (share->version == 0) ? 1 : 0;
|
2001-07-11 09:36:22 +02:00
|
|
|
start_list= &(*start_list)->next;
|
2001-09-27 20:45:48 +02:00
|
|
|
*start_list=0;
|
2000-12-18 22:22:20 +01:00
|
|
|
}
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2001-07-11 09:36:22 +02:00
|
|
|
DBUG_RETURN(open_list);
|
2000-12-18 22:22:20 +01:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
/*****************************************************************************
|
|
|
|
* Functions to free open table cache
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
|
|
|
|
|
|
void intern_close_table(TABLE *table)
|
|
|
|
{ // Free all structures
|
2005-11-23 21:45:02 +01:00
|
|
|
DBUG_ENTER("intern_close_table");
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
DBUG_PRINT("tcache", ("table: '%s'.'%s' 0x%lx",
|
|
|
|
table->s ? table->s->db.str : "?",
|
|
|
|
table->s ? table->s->table_name.str : "?",
|
|
|
|
(long) table));
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
free_io_cache(table);
|
2004-09-07 14:29:46 +02:00
|
|
|
delete table->triggers;
|
2009-12-01 14:51:50 +01:00
|
|
|
if (table->file) // Not true if placeholder
|
2009-11-24 14:54:59 +01:00
|
|
|
(void) closefrm(table, 1); // close file
|
2005-11-23 21:45:02 +01:00
|
|
|
DBUG_VOID_RETURN;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2002-08-05 17:50:38 +02:00
|
|
|
/*
|
|
|
|
Remove table from the open table cache
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
free_cache_entry()
|
|
|
|
table Table to remove
|
|
|
|
|
|
|
|
NOTE
|
|
|
|
We need to have a lock on LOCK_open when calling this
|
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
static void free_cache_entry(TABLE *table)
|
|
|
|
{
|
|
|
|
DBUG_ENTER("free_cache_entry");
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/* This should be done before releasing table share. */
|
|
|
|
table_def_remove_table(table);
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
intern_close_table(table);
|
2009-11-30 16:55:03 +01:00
|
|
|
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
my_free((uchar*) table,MYF(0));
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
}
|
|
|
|
|
2003-01-17 15:33:54 +01:00
|
|
|
/* Free resources allocated by filesort() and read_record() */
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
void free_io_cache(TABLE *table)
|
|
|
|
{
|
2002-01-16 22:02:26 +01:00
|
|
|
DBUG_ENTER("free_io_cache");
|
2003-04-24 13:33:33 +02:00
|
|
|
if (table->sort.io_cache)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2003-04-24 13:33:33 +02:00
|
|
|
close_cached_file(table->sort.io_cache);
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
my_free((uchar*) table->sort.io_cache,MYF(0));
|
2003-04-24 13:33:33 +02:00
|
|
|
table->sort.io_cache=0;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2002-01-16 22:02:26 +01:00
|
|
|
DBUG_VOID_RETURN;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2007-03-24 09:18:19 +01:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/**
|
|
|
|
Auxiliary function which allows to kill delayed threads for
|
|
|
|
particular table identified by its share.
|
|
|
|
|
|
|
|
@param share Table share.
|
2009-12-01 20:13:01 +01:00
|
|
|
|
|
|
|
@pre Caller should have LOCK_open mutex acquired.
|
2009-11-30 16:55:03 +01:00
|
|
|
*/
|
|
|
|
|
|
|
|
static void kill_delayed_threads_for_table(TABLE_SHARE *share)
|
|
|
|
{
|
|
|
|
I_P_List_iterator<TABLE, TABLE_share> it(share->used_tables);
|
|
|
|
TABLE *tab;
|
2009-12-01 20:13:01 +01:00
|
|
|
|
|
|
|
safe_mutex_assert_owner(&LOCK_open);
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
while ((tab= it++))
|
|
|
|
{
|
|
|
|
THD *in_use= tab->in_use;
|
|
|
|
|
|
|
|
if ((in_use->system_thread & SYSTEM_THREAD_DELAYED_INSERT) &&
|
|
|
|
! in_use->killed)
|
|
|
|
{
|
|
|
|
in_use->killed= THD::KILL_CONNECTION;
|
|
|
|
pthread_mutex_lock(&in_use->mysys_var->mutex);
|
|
|
|
if (in_use->mysys_var->current_cond)
|
|
|
|
{
|
|
|
|
pthread_mutex_lock(in_use->mysys_var->current_mutex);
|
|
|
|
pthread_cond_broadcast(in_use->mysys_var->current_cond);
|
|
|
|
pthread_mutex_unlock(in_use->mysys_var->current_mutex);
|
|
|
|
}
|
|
|
|
pthread_mutex_unlock(&in_use->mysys_var->mutex);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-01-24 15:48:25 +01:00
|
|
|
/*
|
|
|
|
Close all tables which aren't in use by any thread
|
|
|
|
|
2007-12-12 22:44:14 +01:00
|
|
|
@param thd Thread context
|
|
|
|
@param tables List of tables to remove from the cache
|
|
|
|
@param have_lock If LOCK_open is locked
|
|
|
|
@param wait_for_refresh Wait for a impending flush
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
@note THD can be NULL, but then wait_for_refresh must be FALSE
|
|
|
|
and tables must be NULL.
|
|
|
|
|
|
|
|
@note When called as part of FLUSH TABLES WITH READ LOCK this function
|
|
|
|
ignores metadata locks held by other threads. In order to avoid
|
|
|
|
situation when FLUSH TABLES WITH READ LOCK sneaks in at the moment
|
|
|
|
when some write-locked table is being reopened (by FLUSH TABLES or
|
|
|
|
ALTER TABLE) we have to rely on additional global shared metadata
|
|
|
|
lock taken by thread trying to obtain global read lock.
|
2005-01-24 15:48:25 +01:00
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2007-12-12 22:44:14 +01:00
|
|
|
bool close_cached_tables(THD *thd, TABLE_LIST *tables, bool have_lock,
|
2009-11-30 16:55:03 +01:00
|
|
|
bool wait_for_refresh)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
bool result= FALSE;
|
|
|
|
bool found= TRUE;
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("close_cached_tables");
|
2007-12-12 22:44:14 +01:00
|
|
|
DBUG_ASSERT(thd || (!wait_for_refresh && !tables));
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2006-01-12 19:51:02 +01:00
|
|
|
if (!have_lock)
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_lock(&LOCK_open);
|
2000-08-21 23:18:32 +02:00
|
|
|
if (!tables)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
refresh_version++; // Force close of open tables
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
DBUG_PRINT("tcache", ("incremented global refresh_version to: %lu",
|
|
|
|
refresh_version));
|
2009-11-30 16:55:03 +01:00
|
|
|
kill_delayed_threads();
|
2009-12-01 20:13:01 +01:00
|
|
|
/*
|
|
|
|
Get rid of all unused TABLE and TABLE_SHARE instances. By doing
|
|
|
|
this we automatically close all tables which were marked as "old".
|
|
|
|
*/
|
|
|
|
while (unused_tables)
|
|
|
|
free_cache_entry(unused_tables);
|
|
|
|
/* Free table shares which were not freed implicitly by loop above. */
|
|
|
|
while (oldest_unused_share->next)
|
|
|
|
(void) my_hash_delete(&table_def_cache, (uchar*) oldest_unused_share);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2000-08-21 23:18:32 +02:00
|
|
|
else
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2000-08-21 23:18:32 +02:00
|
|
|
bool found=0;
|
2004-07-16 00:15:55 +02:00
|
|
|
for (TABLE_LIST *table= tables; table; table= table->next_local)
|
2000-08-21 23:18:32 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
TABLE_SHARE *share= get_cached_table_share(table->db, table->table_name);
|
|
|
|
|
|
|
|
if (share)
|
|
|
|
{
|
|
|
|
kill_delayed_threads_for_table(share);
|
2009-12-01 20:13:01 +01:00
|
|
|
/* tdc_remove_table() also sets TABLE_SHARE::version to 0. */
|
|
|
|
tdc_remove_table(thd, TDC_RT_REMOVE_UNUSED, table->db,
|
|
|
|
table->table_name);
|
2000-08-21 23:18:32 +02:00
|
|
|
found=1;
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
2000-08-21 23:18:32 +02:00
|
|
|
}
|
|
|
|
if (!found)
|
2007-12-12 22:44:14 +01:00
|
|
|
wait_for_refresh=0; // Nothing to wait for
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
|
2009-12-01 20:13:01 +01:00
|
|
|
if (!have_lock)
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
if (!wait_for_refresh)
|
|
|
|
DBUG_RETURN(result);
|
|
|
|
|
2009-12-01 20:13:01 +01:00
|
|
|
/* Code below assume that LOCK_open is released. */
|
2009-11-30 16:55:03 +01:00
|
|
|
DBUG_ASSERT(!have_lock);
|
|
|
|
|
2009-12-01 15:39:03 +01:00
|
|
|
if (thd->locked_tables_mode)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
/*
|
2009-12-02 16:22:15 +01:00
|
|
|
If we are under LOCK TABLES, we need to reopen the tables without
|
2009-11-30 16:55:03 +01:00
|
|
|
opening a door for any concurrent threads to sneak in and get
|
|
|
|
lock on our tables. To achieve this we use exclusive metadata
|
|
|
|
locks.
|
2000-07-31 21:29:14 +02:00
|
|
|
*/
|
2009-12-02 16:22:15 +01:00
|
|
|
TABLE_LIST *tables_to_reopen= (tables ? tables :
|
|
|
|
thd->locked_tables_list.locked_tables());
|
|
|
|
|
|
|
|
for (TABLE_LIST *table_list= tables_to_reopen; table_list;
|
|
|
|
table_list= table_list->next_global)
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
/* A check that the table was locked for write is done by the caller. */
|
2009-12-09 10:44:01 +01:00
|
|
|
TABLE *table= find_table_for_mdl_upgrade(thd->open_tables, table_list->db,
|
|
|
|
table_list->table_name);
|
2009-12-02 16:22:15 +01:00
|
|
|
|
|
|
|
/* May return NULL if this table has already been closed via an alias. */
|
|
|
|
if (! table)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (wait_while_table_is_used(thd, table, HA_EXTRA_FORCE_REOPEN))
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
result= TRUE;
|
|
|
|
goto err_with_reopen;
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
2009-12-02 16:22:15 +01:00
|
|
|
close_all_tables_for_name(thd, table->s, FALSE);
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Wait until all threads have closed all the tables we are flushing. */
|
|
|
|
DBUG_PRINT("info", ("Waiting for other threads to close their open tables"));
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
while (found && ! thd->killed)
|
|
|
|
{
|
|
|
|
found= FALSE;
|
|
|
|
/*
|
|
|
|
To avoid self and other kinds of deadlock we have to flush open HANDLERs.
|
|
|
|
*/
|
2007-11-20 18:17:53 +01:00
|
|
|
mysql_ha_flush(thd);
|
2009-09-29 17:38:40 +02:00
|
|
|
DEBUG_SYNC(thd, "after_flush_unlock");
|
2007-11-20 18:17:53 +01:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
|
|
|
|
thd->enter_cond(&COND_refresh, &LOCK_open, "Flushing tables");
|
|
|
|
|
|
|
|
if (!tables)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
for (uint idx=0 ; idx < table_def_cache.records ; idx++)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
TABLE_SHARE *share=(TABLE_SHARE*) my_hash_element(&table_def_cache,
|
|
|
|
idx);
|
|
|
|
if (share->version != refresh_version)
|
|
|
|
{
|
|
|
|
found= TRUE;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
for (TABLE_LIST *table= tables; table; table= table->next_local)
|
|
|
|
{
|
|
|
|
TABLE_SHARE *share= get_cached_table_share(table->db, table->table_name);
|
|
|
|
if (share && share->version != refresh_version)
|
|
|
|
{
|
|
|
|
found= TRUE;
|
|
|
|
break;
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
if (found)
|
|
|
|
{
|
|
|
|
DBUG_PRINT("signal", ("Waiting for COND_refresh"));
|
|
|
|
pthread_cond_wait(&COND_refresh,&LOCK_open);
|
|
|
|
}
|
|
|
|
|
|
|
|
thd->exit_cond(NULL);
|
|
|
|
}
|
|
|
|
|
|
|
|
err_with_reopen:
|
2009-12-01 15:39:03 +01:00
|
|
|
if (thd->locked_tables_mode)
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
2000-07-31 21:29:14 +02:00
|
|
|
/*
|
|
|
|
No other thread has the locked tables open; reopen them and get the
|
|
|
|
old locks. This should always succeed (unless some external process
|
|
|
|
has removed the tables)
|
|
|
|
*/
|
2009-12-02 16:22:15 +01:00
|
|
|
thd->locked_tables_list.reopen_tables(thd);
|
2009-11-30 23:39:13 +01:00
|
|
|
/*
|
2009-12-04 00:52:05 +01:00
|
|
|
Since downgrade_exclusive_lock() won't do anything with shared
|
|
|
|
metadata lock it is much simpler to go through all open tables rather
|
2009-11-30 23:39:13 +01:00
|
|
|
than picking only those tables that were flushed.
|
|
|
|
*/
|
|
|
|
for (TABLE *tab= thd->open_tables; tab; tab= tab->next)
|
2009-12-04 00:52:05 +01:00
|
|
|
tab->mdl_ticket->downgrade_exclusive_lock();
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
DBUG_RETURN(result);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-03-24 09:18:19 +01:00
|
|
|
/*
|
|
|
|
Close all tables which match specified connection string or
|
|
|
|
if specified string is NULL, then any table with a connection string.
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool close_cached_connection_tables(THD *thd, bool if_wait_for_refresh,
|
|
|
|
LEX_STRING *connection, bool have_lock)
|
|
|
|
{
|
|
|
|
uint idx;
|
|
|
|
TABLE_LIST tmp, *tables= NULL;
|
|
|
|
bool result= FALSE;
|
|
|
|
DBUG_ENTER("close_cached_connections");
|
|
|
|
DBUG_ASSERT(thd);
|
|
|
|
|
|
|
|
bzero(&tmp, sizeof(TABLE_LIST));
|
2007-12-12 22:44:14 +01:00
|
|
|
|
2007-03-24 09:18:19 +01:00
|
|
|
if (!have_lock)
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_lock(&LOCK_open);
|
2007-12-12 22:44:14 +01:00
|
|
|
|
2007-03-24 09:18:19 +01:00
|
|
|
for (idx= 0; idx < table_def_cache.records; idx++)
|
|
|
|
{
|
2009-10-14 18:37:38 +02:00
|
|
|
TABLE_SHARE *share= (TABLE_SHARE *) my_hash_element(&table_def_cache, idx);
|
2007-03-24 09:18:19 +01:00
|
|
|
|
|
|
|
/* Ignore if table is not open or does not have a connect_string */
|
|
|
|
if (!share->connect_string.length || !share->ref_count)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
/* Compare the connection string */
|
|
|
|
if (connection &&
|
|
|
|
(connection->length > share->connect_string.length ||
|
|
|
|
(connection->length < share->connect_string.length &&
|
|
|
|
(share->connect_string.str[connection->length] != '/' &&
|
|
|
|
share->connect_string.str[connection->length] != '\\')) ||
|
|
|
|
strncasecmp(connection->str, share->connect_string.str,
|
|
|
|
connection->length)))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
/* close_cached_tables() only uses these elements */
|
|
|
|
tmp.db= share->db.str;
|
|
|
|
tmp.table_name= share->table_name.str;
|
|
|
|
tmp.next_local= tables;
|
|
|
|
|
|
|
|
tables= (TABLE_LIST *) memdup_root(thd->mem_root, (char*)&tmp,
|
|
|
|
sizeof(TABLE_LIST));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (tables)
|
2009-11-30 16:55:03 +01:00
|
|
|
result= close_cached_tables(thd, tables, TRUE, FALSE);
|
2007-12-12 22:44:14 +01:00
|
|
|
|
2007-03-24 09:18:19 +01:00
|
|
|
if (!have_lock)
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2007-12-12 22:44:14 +01:00
|
|
|
|
2007-03-24 09:18:19 +01:00
|
|
|
if (if_wait_for_refresh)
|
|
|
|
{
|
|
|
|
pthread_mutex_lock(&thd->mysys_var->mutex);
|
|
|
|
thd->mysys_var->current_mutex= 0;
|
|
|
|
thd->mysys_var->current_cond= 0;
|
|
|
|
thd->proc_info=0;
|
|
|
|
pthread_mutex_unlock(&thd->mysys_var->mutex);
|
|
|
|
}
|
|
|
|
|
|
|
|
DBUG_RETURN(result);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-11-01 21:52:56 +01:00
|
|
|
/**
|
|
|
|
Mark all temporary tables which were used by the current statement or
|
|
|
|
substatement as free for reuse, but only if the query_id can be cleared.
|
|
|
|
|
|
|
|
@param thd thread context
|
|
|
|
|
|
|
|
@remark For temp tables associated with a open SQL HANDLER the query_id
|
|
|
|
is not reset until the HANDLER is closed.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void mark_temp_tables_as_free_for_reuse(THD *thd)
|
|
|
|
{
|
|
|
|
for (TABLE *table= thd->temporary_tables ; table ; table= table->next)
|
|
|
|
{
|
|
|
|
if ((table->query_id == thd->query_id) && ! table->open_by_handler)
|
|
|
|
{
|
|
|
|
table->query_id= 0;
|
|
|
|
table->file->ha_reset();
|
2009-12-03 00:09:22 +01:00
|
|
|
|
|
|
|
/* Detach temporary MERGE children from temporary parent. */
|
|
|
|
DBUG_ASSERT(table->file);
|
|
|
|
table->file->extra(HA_EXTRA_DETACH_CHILDREN);
|
|
|
|
|
2009-01-07 13:11:37 +01:00
|
|
|
/*
|
|
|
|
Reset temporary table lock type to it's default value (TL_WRITE).
|
|
|
|
|
|
|
|
Statements such as INSERT INTO .. SELECT FROM tmp, CREATE TABLE
|
|
|
|
.. SELECT FROM tmp and UPDATE may under some circumstances modify
|
|
|
|
the lock type of the tables participating in the statement. This
|
|
|
|
isn't a problem for non-temporary tables since their lock type is
|
|
|
|
reset at every open, but the same does not occur for temporary
|
|
|
|
tables for historical reasons.
|
|
|
|
|
|
|
|
Furthermore, the lock type of temporary tables is not really that
|
|
|
|
important because they can only be used by one query at a time and
|
|
|
|
not even twice in a query -- a temporary table is represented by
|
|
|
|
only one TABLE object. Nonetheless, it's safer from a maintenance
|
|
|
|
point of view to reset the lock type of this singleton TABLE object
|
|
|
|
as to not cause problems when the table is reused.
|
|
|
|
|
|
|
|
Even under LOCK TABLES mode its okay to reset the lock type as
|
|
|
|
LOCK TABLES is allowed (but ignored) for a temporary table.
|
|
|
|
*/
|
|
|
|
table->reginfo.lock_type= TL_WRITE;
|
2007-11-01 21:52:56 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2003-01-07 10:45:06 +01:00
|
|
|
/*
|
2005-03-04 14:35:28 +01:00
|
|
|
Mark all tables in the list which were used by current substatement
|
|
|
|
as free for reuse.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
mark_used_tables_as_free_for_reuse()
|
|
|
|
thd - thread context
|
|
|
|
table - head of the list of tables
|
|
|
|
|
|
|
|
DESCRIPTION
|
|
|
|
Marks all tables in the list which were used by current substatement
|
|
|
|
(they are marked by its query_id) as free for reuse.
|
2006-02-24 17:34:15 +01:00
|
|
|
|
|
|
|
NOTE
|
|
|
|
The reason we reset query_id is that it's not enough to just test
|
|
|
|
if table->query_id != thd->query_id to know if a table is in use.
|
|
|
|
|
|
|
|
For example
|
|
|
|
SELECT f1_that_uses_t1() FROM t1;
|
|
|
|
In f1_that_uses_t1() we will see one instance of t1 where query_id is
|
|
|
|
set to query_id of original query.
|
2005-03-04 14:35:28 +01:00
|
|
|
*/
|
|
|
|
|
|
|
|
static void mark_used_tables_as_free_for_reuse(THD *thd, TABLE *table)
|
|
|
|
{
|
|
|
|
for (; table ; table= table->next)
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
DBUG_ASSERT(table->pos_in_locked_tables == NULL ||
|
|
|
|
table->pos_in_locked_tables->table == table);
|
2005-03-04 14:35:28 +01:00
|
|
|
if (table->query_id == thd->query_id)
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
{
|
2005-03-04 14:35:28 +01:00
|
|
|
table->query_id= 0;
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
table->file->ha_reset();
|
|
|
|
}
|
|
|
|
}
|
2005-03-04 14:35:28 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-11-01 21:52:56 +01:00
|
|
|
/**
|
|
|
|
Auxiliary function to close all tables in the open_tables list.
|
|
|
|
|
|
|
|
@param thd Thread context.
|
|
|
|
|
|
|
|
@remark It should not ordinarily be called directly.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void close_open_tables(THD *thd)
|
|
|
|
{
|
|
|
|
bool found_old_table= 0;
|
|
|
|
|
|
|
|
safe_mutex_assert_not_owner(&LOCK_open);
|
|
|
|
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_lock(&LOCK_open);
|
2007-11-01 21:52:56 +01:00
|
|
|
|
|
|
|
DBUG_PRINT("info", ("thd->open_tables: 0x%lx", (long) thd->open_tables));
|
|
|
|
|
|
|
|
while (thd->open_tables)
|
|
|
|
found_old_table|= close_thread_table(thd, &thd->open_tables);
|
|
|
|
thd->some_tables_deleted= 0;
|
|
|
|
|
|
|
|
/* Free tables to hold down open files */
|
2009-11-30 16:55:03 +01:00
|
|
|
while (table_cache_count > table_cache_size && unused_tables)
|
|
|
|
free_cache_entry(unused_tables);
|
2007-11-01 21:52:56 +01:00
|
|
|
if (found_old_table)
|
|
|
|
{
|
|
|
|
/* Tell threads waiting for refresh that something has happened */
|
|
|
|
broadcast_refresh();
|
|
|
|
}
|
|
|
|
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2007-11-01 21:52:56 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
/**
|
|
|
|
Close all open instances of the table but keep the MDL lock,
|
|
|
|
if any.
|
|
|
|
|
|
|
|
Works both under LOCK TABLES and in the normal mode.
|
|
|
|
Removes all closed instances of the table from the table cache.
|
|
|
|
|
|
|
|
@param thd thread handle
|
|
|
|
@param[in] share table share, but is just a handy way to
|
|
|
|
access the table cache key
|
|
|
|
|
|
|
|
@param[in] remove_from_locked_tables
|
|
|
|
TRUE if the table is being dropped or renamed.
|
|
|
|
In that case the documented behaviour is to
|
|
|
|
implicitly remove the table from LOCK TABLES
|
|
|
|
list.
|
|
|
|
*/
|
|
|
|
|
|
|
|
void
|
|
|
|
close_all_tables_for_name(THD *thd, TABLE_SHARE *share,
|
|
|
|
bool remove_from_locked_tables)
|
|
|
|
{
|
|
|
|
char key[MAX_DBKEY_LENGTH];
|
|
|
|
uint key_length= share->table_cache_key.length;
|
|
|
|
|
|
|
|
memcpy(key, share->table_cache_key.str, key_length);
|
|
|
|
|
|
|
|
safe_mutex_assert_not_owner(&LOCK_open);
|
|
|
|
/*
|
|
|
|
We need to hold LOCK_open while changing the open_tables
|
|
|
|
list, since another thread may work on it.
|
2009-12-02 17:31:57 +01:00
|
|
|
@sa mysql_notify_thread_having_shared_lock()
|
2009-12-02 16:22:15 +01:00
|
|
|
*/
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
|
|
|
|
for (TABLE **prev= &thd->open_tables; *prev; )
|
|
|
|
{
|
|
|
|
TABLE *table= *prev;
|
|
|
|
|
|
|
|
if (table->s->table_cache_key.length == key_length &&
|
|
|
|
!memcmp(table->s->table_cache_key.str, key, key_length))
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
Does nothing if the table is not locked.
|
|
|
|
This allows one to use this function after a table
|
|
|
|
has been unlocked, e.g. in partition management.
|
|
|
|
*/
|
|
|
|
mysql_lock_remove(thd, thd->lock, table);
|
|
|
|
|
|
|
|
thd->locked_tables_list.unlink_from_list(thd,
|
|
|
|
table->pos_in_locked_tables,
|
|
|
|
remove_from_locked_tables);
|
|
|
|
|
|
|
|
/* Make sure the table is removed from the cache */
|
|
|
|
table->s->version= 0;
|
|
|
|
close_thread_table(thd, prev);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/* Step to next entry in open_tables list. */
|
|
|
|
prev= &table->next;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
/* We have been removing tables from the table cache. */
|
|
|
|
broadcast_refresh();
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-03-04 14:35:28 +01:00
|
|
|
/*
|
|
|
|
Close all tables used by the current substatement, or all tables
|
|
|
|
used by this thread if we are on the upper level.
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2003-01-07 10:45:06 +01:00
|
|
|
SYNOPSIS
|
|
|
|
close_thread_tables()
|
|
|
|
thd Thread handler
|
|
|
|
|
|
|
|
IMPLEMENTATION
|
|
|
|
Unlocks tables and frees derived tables.
|
|
|
|
Put all normal tables used by thread in free list.
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2007-11-01 21:52:56 +01:00
|
|
|
It will only close/mark as free for reuse tables opened by this
|
|
|
|
substatement, it will also check if we are closing tables after
|
|
|
|
execution of complete query (i.e. we are on upper level) and will
|
|
|
|
leave prelocked mode if needed.
|
2003-01-07 10:45:06 +01:00
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
void close_thread_tables(THD *thd)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
TABLE *table;
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("close_thread_tables");
|
|
|
|
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
#ifdef EXTRA_DEBUG
|
|
|
|
DBUG_PRINT("tcache", ("open tables:"));
|
|
|
|
for (table= thd->open_tables; table; table= table->next)
|
|
|
|
DBUG_PRINT("tcache", ("table: '%s'.'%s' 0x%lx", table->s->db.str,
|
|
|
|
table->s->table_name.str, (long) table));
|
|
|
|
#endif
|
|
|
|
|
2009-12-03 00:09:22 +01:00
|
|
|
/* Detach MERGE children after every statement. Even under LOCK TABLES. */
|
|
|
|
for (table= thd->open_tables; table; table= table->next)
|
|
|
|
{
|
|
|
|
/* Table might be in use by some outer statement. */
|
|
|
|
DBUG_PRINT("tcache", ("table: '%s' query_id: %lu",
|
|
|
|
table->s->table_name.str, (ulong) table->query_id));
|
|
|
|
if (thd->locked_tables_mode <= LTM_LOCK_TABLES ||
|
|
|
|
table->query_id == thd->query_id)
|
|
|
|
{
|
|
|
|
DBUG_ASSERT(table->file);
|
|
|
|
table->file->extra(HA_EXTRA_DETACH_CHILDREN);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2005-03-04 14:35:28 +01:00
|
|
|
/*
|
|
|
|
We are assuming here that thd->derived_tables contains ONLY derived
|
|
|
|
tables for this substatement. i.e. instead of approach which uses
|
|
|
|
query_id matching for determining which of the derived tables belong
|
|
|
|
to this substatement we rely on the ability of substatements to
|
|
|
|
save/restore thd->derived_tables during their execution.
|
|
|
|
|
|
|
|
TODO: Probably even better approach is to simply associate list of
|
|
|
|
derived tables with (sub-)statement instead of thread and destroy
|
|
|
|
them at the end of its execution.
|
|
|
|
*/
|
2007-11-01 21:52:56 +01:00
|
|
|
if (thd->derived_tables)
|
2003-01-07 10:45:06 +01:00
|
|
|
{
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
TABLE *next;
|
2003-01-07 10:45:06 +01:00
|
|
|
/*
|
2005-01-16 13:16:23 +01:00
|
|
|
Close all derived tables generated in queries like
|
|
|
|
SELECT * FROM (SELECT * FROM t1)
|
2003-01-07 10:45:06 +01:00
|
|
|
*/
|
|
|
|
for (table= thd->derived_tables ; table ; table= next)
|
|
|
|
{
|
|
|
|
next= table->next;
|
|
|
|
free_tmp_table(thd, table);
|
|
|
|
}
|
|
|
|
thd->derived_tables= 0;
|
|
|
|
}
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2007-11-01 21:52:56 +01:00
|
|
|
/*
|
|
|
|
Mark all temporary tables used by this statement as free for reuse.
|
|
|
|
*/
|
|
|
|
mark_temp_tables_as_free_for_reuse(thd);
|
2008-02-19 12:43:01 +01:00
|
|
|
/*
|
|
|
|
Let us commit transaction for statement. Since in 5.0 we only have
|
|
|
|
one statement transaction and don't allow several nested statement
|
|
|
|
transactions this call will do nothing if we are inside of stored
|
|
|
|
function or trigger (i.e. statement transaction is already active and
|
|
|
|
does not belong to statement for which we do close_thread_tables()).
|
|
|
|
TODO: This should be fixed in later releases.
|
|
|
|
*/
|
|
|
|
if (!(thd->state_flags & Open_tables_state::BACKUPS_AVAIL))
|
2005-03-04 14:35:28 +01:00
|
|
|
{
|
2009-09-10 11:18:29 +02:00
|
|
|
thd->stmt_da->can_overwrite_status= TRUE;
|
2009-12-03 19:37:38 +01:00
|
|
|
thd->is_error() ? trans_rollback_stmt(thd) : trans_commit_stmt(thd);
|
2009-09-10 11:18:29 +02:00
|
|
|
thd->stmt_da->can_overwrite_status= FALSE;
|
2008-02-19 12:43:01 +01:00
|
|
|
|
2005-03-04 14:35:28 +01:00
|
|
|
/*
|
2008-02-19 12:43:01 +01:00
|
|
|
Reset transaction state, but only if we're not inside a
|
|
|
|
sub-statement of a prelocked statement.
|
2005-03-04 14:35:28 +01:00
|
|
|
*/
|
2009-12-01 15:39:03 +01:00
|
|
|
if (thd->locked_tables_mode <= LTM_LOCK_TABLES ||
|
|
|
|
thd->lex->requires_prelocking())
|
2008-02-19 12:43:01 +01:00
|
|
|
thd->transaction.stmt.reset();
|
|
|
|
}
|
|
|
|
|
2009-12-01 15:39:03 +01:00
|
|
|
if (thd->locked_tables_mode)
|
2008-02-19 12:43:01 +01:00
|
|
|
{
|
2005-03-04 14:35:28 +01:00
|
|
|
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
/* Ensure we are calling ha_reset() for all used tables */
|
|
|
|
mark_used_tables_as_free_for_reuse(thd, thd->open_tables);
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2008-02-19 12:43:01 +01:00
|
|
|
/*
|
|
|
|
We are under simple LOCK TABLES or we're inside a sub-statement
|
|
|
|
of a prelocked statement, so should not do anything else.
|
2009-12-01 16:04:32 +01:00
|
|
|
|
|
|
|
Note that even if we are in LTM_LOCK_TABLES mode and statement
|
|
|
|
requires prelocking (e.g. when we are closing tables after
|
|
|
|
failing ot "open" all tables required for statement execution)
|
|
|
|
we will exit this function a few lines below.
|
2008-02-19 12:43:01 +01:00
|
|
|
*/
|
2009-12-01 15:39:03 +01:00
|
|
|
if (! thd->lex->requires_prelocking())
|
2005-03-04 14:35:28 +01:00
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
|
|
|
|
/*
|
2008-02-19 12:43:01 +01:00
|
|
|
We are in the top-level statement of a prelocked statement,
|
|
|
|
so we have to leave the prelocked mode now with doing implicit
|
|
|
|
UNLOCK TABLES if needed.
|
2005-03-04 14:35:28 +01:00
|
|
|
*/
|
2009-12-01 15:39:03 +01:00
|
|
|
if (thd->locked_tables_mode == LTM_PRELOCKED_UNDER_LOCK_TABLES)
|
|
|
|
thd->locked_tables_mode= LTM_LOCK_TABLES;
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2009-12-01 15:39:03 +01:00
|
|
|
if (thd->locked_tables_mode == LTM_LOCK_TABLES)
|
2005-03-04 14:35:28 +01:00
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
|
2009-12-01 15:39:03 +01:00
|
|
|
thd->locked_tables_mode= LTM_NONE;
|
|
|
|
|
2005-03-04 14:35:28 +01:00
|
|
|
/* Fallthrough */
|
2001-09-06 22:57:32 +02:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
if (thd->lock)
|
|
|
|
{
|
2006-02-16 08:30:53 +01:00
|
|
|
/*
|
|
|
|
For RBR we flush the pending event just before we unlock all the
|
|
|
|
tables. This means that we are at the end of a topmost
|
|
|
|
statement, so we ensure that the STMT_END_F flag is set on the
|
|
|
|
pending event. For statements that are *inside* stored
|
|
|
|
functions, the pending event will not be flushed: that will be
|
|
|
|
handled either before writing a query log event (inside
|
|
|
|
binlog_query()) or when preparing a pending event.
|
|
|
|
*/
|
2006-05-05 19:08:40 +02:00
|
|
|
thd->binlog_flush_pending_rows_event(TRUE);
|
2003-01-07 10:45:06 +01:00
|
|
|
mysql_unlock_tables(thd, thd->lock);
|
|
|
|
thd->lock=0;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
/*
|
|
|
|
Note that we need to hold LOCK_open while changing the
|
|
|
|
open_tables list. Another thread may work on it.
|
2009-12-02 17:31:57 +01:00
|
|
|
(See: mysql_notify_thread_having_shared_lock())
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
Closing a MERGE child before the parent would be fatal if the
|
|
|
|
other thread tries to abort the MERGE lock in between.
|
|
|
|
*/
|
2007-11-01 21:52:56 +01:00
|
|
|
if (thd->open_tables)
|
|
|
|
close_open_tables(thd);
|
2005-03-04 14:35:28 +01:00
|
|
|
|
Backport of revno ## 2617.31.1, 2617.31.3, 2617.31.4, 2617.31.5,
2617.31.12, 2617.31.15, 2617.31.15, 2617.31.16, 2617.43.1
- initial changeset that introduced the fix for
Bug#989 and follow up fixes for all test suite failures
introduced in the initial changeset.
------------------------------------------------------------
revno: 2617.31.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Fri 2009-03-06 19:17:00 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
Currently the MySQL server does not keep metadata locks on
schema objects for the duration of a transaction, thus failing
to guarantee the integrity of the schema objects being used
during the transaction and to protect then from concurrent
DDL operations. This also poses a problem for replication as
a DDL operation might be replicated even thought there are
active transactions using the object being modified.
The solution is to defer the release of metadata locks until
a active transaction is either committed or rolled back. This
prevents other statements from modifying the table for the
entire duration of the transaction. This provides commitment
ordering for guaranteeing serializability across multiple
transactions.
- Incompatible change:
If MySQL's metadata locking system encounters a lock conflict,
the usual schema is to use the try and back-off technique to
avoid deadlocks -- this schema consists in releasing all locks
and trying to acquire them all in one go.
But in a transactional context this algorithm can't be utilized
as its not possible to release locks acquired during the course
of the transaction without breaking the transaction commitments.
To avoid deadlocks in this case, the ER_LOCK_DEADLOCK will be
returned if a lock conflict is encountered during a transaction.
Let's consider an example:
A transaction has two statements that modify table t1, then table
t2, and then commits. The first statement of the transaction will
acquire a shared metadata lock on table t1, and it will be kept
utill COMMIT to ensure serializability.
At the moment when the second statement attempts to acquire a
shared metadata lock on t2, a concurrent ALTER or DROP statement
might have locked t2 exclusively. The prescription of the current
locking protocol is that the acquirer of the shared lock backs off
-- gives up all his current locks and retries. This implies that
the entire multi-statement transaction has to be rolled back.
- Incompatible change:
FLUSH commands such as FLUSH PRIVILEGES and FLUSH TABLES WITH READ
LOCK won't cause locked tables to be implicitly unlocked anymore.
2009-12-05 00:02:48 +01:00
|
|
|
/*
|
|
|
|
Defer the release of metadata locks until the current transaction
|
|
|
|
is either committed or rolled back. This prevents other statements
|
|
|
|
from modifying the table for the entire duration of this transaction.
|
|
|
|
This provides commitment ordering for guaranteeing serializability
|
|
|
|
across multiple transactions.
|
|
|
|
*/
|
|
|
|
if (!thd->in_multi_stmt_transaction() ||
|
|
|
|
(thd->state_flags & Open_tables_state::BACKUPS_AVAIL))
|
|
|
|
thd->mdl_context.release_all_locks();
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
}
|
|
|
|
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
|
2001-07-11 09:36:22 +02:00
|
|
|
/* move one table to free list */
|
|
|
|
|
|
|
|
bool close_thread_table(THD *thd, TABLE **table_ptr)
|
|
|
|
{
|
2003-05-03 15:21:43 +02:00
|
|
|
bool found_old_table= 0;
|
|
|
|
TABLE *table= *table_ptr;
|
2005-07-13 11:48:13 +02:00
|
|
|
DBUG_ENTER("close_thread_table");
|
2003-05-03 15:21:43 +02:00
|
|
|
DBUG_ASSERT(table->key_read == 0);
|
2007-05-11 19:51:03 +02:00
|
|
|
DBUG_ASSERT(!table->file || table->file->inited == handler::NONE);
|
2009-11-30 16:55:03 +01:00
|
|
|
safe_mutex_assert_owner(&LOCK_open);
|
2001-07-11 09:36:22 +02:00
|
|
|
|
|
|
|
*table_ptr=table->next;
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
|
2009-12-04 00:52:05 +01:00
|
|
|
table->mdl_ticket= NULL;
|
2009-11-30 23:01:27 +01:00
|
|
|
if (table->needs_reopen() ||
|
2009-12-08 09:26:49 +01:00
|
|
|
thd->version != refresh_version || !table->db_stat ||
|
|
|
|
table_def_shutdown_in_progress)
|
2001-07-11 09:36:22 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
free_cache_entry(table);
|
2001-07-11 09:36:22 +02:00
|
|
|
found_old_table=1;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2009-12-03 00:09:22 +01:00
|
|
|
/* Avoid to have MERGE tables with attached children in unused_tables. */
|
|
|
|
DBUG_ASSERT(table->file);
|
|
|
|
table->file->extra(HA_EXTRA_DETACH_CHILDREN);
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
|
2007-02-28 22:25:50 +01:00
|
|
|
/* Free memory and reset for next loop */
|
2008-07-24 22:38:44 +02:00
|
|
|
free_field_buffers_larger_than(table,MAX_TDC_BLOB_SIZE);
|
|
|
|
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
table->file->ha_reset();
|
2009-11-30 16:55:03 +01:00
|
|
|
table_def_unuse_table(table);
|
2001-07-11 09:36:22 +02:00
|
|
|
}
|
|
|
|
DBUG_RETURN(found_old_table);
|
|
|
|
}
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2006-04-25 21:53:50 +02:00
|
|
|
/* close_temporary_tables' internal, 4 is due to uint4korr definition */
|
|
|
|
static inline uint tmpkeyval(THD *thd, TABLE *table)
|
|
|
|
{
|
|
|
|
return uint4korr(table->s->table_cache_key.str + table->s->table_cache_key.length - 4);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
/*
|
|
|
|
Close all temporary tables created by 'CREATE TEMPORARY TABLE' for thread
|
2006-04-25 21:53:50 +02:00
|
|
|
creates one DROP TEMPORARY TABLE binlog event for each pseudo-thread
|
2005-11-23 21:45:02 +01:00
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
void close_temporary_tables(THD *thd)
|
|
|
|
{
|
2006-05-29 17:01:58 +02:00
|
|
|
TABLE *table;
|
2009-06-17 16:56:44 +02:00
|
|
|
TABLE *next= NULL;
|
2006-11-27 00:47:38 +01:00
|
|
|
TABLE *prev_table;
|
|
|
|
/* Assume thd->options has OPTION_QUOTE_SHOW_CREATE */
|
|
|
|
bool was_quote_show= TRUE;
|
|
|
|
|
2002-11-07 03:02:37 +01:00
|
|
|
if (!thd->temporary_tables)
|
|
|
|
return;
|
2000-12-07 14:48:05 +01:00
|
|
|
|
2009-02-23 04:26:38 +01:00
|
|
|
if (!mysql_bin_log.is_open() ||
|
|
|
|
(thd->current_stmt_binlog_row_based && thd->variables.binlog_format == BINLOG_FORMAT_ROW))
|
2006-05-29 17:01:58 +02:00
|
|
|
{
|
2007-08-03 00:14:27 +02:00
|
|
|
TABLE *tmp_next;
|
|
|
|
for (table= thd->temporary_tables; table; table= tmp_next)
|
2006-05-29 17:01:58 +02:00
|
|
|
{
|
2007-08-03 00:14:27 +02:00
|
|
|
tmp_next= table->next;
|
2006-05-29 17:01:58 +02:00
|
|
|
close_temporary(table, 1, 1);
|
|
|
|
}
|
|
|
|
thd->temporary_tables= 0;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2006-11-27 00:47:38 +01:00
|
|
|
/* Better add "if exists", in case a RESET MASTER has been done */
|
2006-05-29 17:01:58 +02:00
|
|
|
const char stub[]= "DROP /*!40005 TEMPORARY */ TABLE IF EXISTS ";
|
|
|
|
uint stub_len= sizeof(stub) - 1;
|
|
|
|
char buf[256];
|
|
|
|
String s_query= String(buf, sizeof(buf), system_charset_info);
|
2007-02-23 12:13:55 +01:00
|
|
|
bool found_user_tables= FALSE;
|
2006-05-29 17:01:58 +02:00
|
|
|
|
2007-02-23 12:13:55 +01:00
|
|
|
memcpy(buf, stub, stub_len);
|
|
|
|
|
2006-05-29 17:01:58 +02:00
|
|
|
/*
|
2007-08-03 00:14:27 +02:00
|
|
|
Insertion sort of temp tables by pseudo_thread_id to build ordered list
|
|
|
|
of sublists of equal pseudo_thread_id
|
2006-04-25 21:53:50 +02:00
|
|
|
*/
|
2006-05-29 17:01:58 +02:00
|
|
|
|
|
|
|
for (prev_table= thd->temporary_tables, table= prev_table->next;
|
2006-04-25 21:53:50 +02:00
|
|
|
table;
|
|
|
|
prev_table= table, table= table->next)
|
2005-11-23 21:45:02 +01:00
|
|
|
{
|
2006-05-29 17:01:58 +02:00
|
|
|
TABLE *prev_sorted /* same as for prev_table */, *sorted;
|
|
|
|
if (is_user_table(table))
|
2006-04-25 21:53:50 +02:00
|
|
|
{
|
2006-05-29 17:01:58 +02:00
|
|
|
if (!found_user_tables)
|
|
|
|
found_user_tables= true;
|
|
|
|
for (prev_sorted= NULL, sorted= thd->temporary_tables; sorted != table;
|
|
|
|
prev_sorted= sorted, sorted= sorted->next)
|
2006-04-25 21:53:50 +02:00
|
|
|
{
|
2006-05-29 17:01:58 +02:00
|
|
|
if (!is_user_table(sorted) ||
|
|
|
|
tmpkeyval(thd, sorted) > tmpkeyval(thd, table))
|
2006-04-25 21:53:50 +02:00
|
|
|
{
|
2006-05-29 17:01:58 +02:00
|
|
|
/* move into the sorted part of the list from the unsorted */
|
|
|
|
prev_table->next= table->next;
|
|
|
|
table->next= sorted;
|
|
|
|
if (prev_sorted)
|
|
|
|
{
|
|
|
|
prev_sorted->next= table;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
thd->temporary_tables= table;
|
|
|
|
}
|
|
|
|
table= prev_table;
|
|
|
|
break;
|
2006-04-25 21:53:50 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
2006-05-29 17:01:58 +02:00
|
|
|
|
|
|
|
/* We always quote db,table names though it is slight overkill */
|
|
|
|
if (found_user_tables &&
|
2006-11-30 02:40:42 +01:00
|
|
|
!(was_quote_show= test(thd->options & OPTION_QUOTE_SHOW_CREATE)))
|
2006-04-25 21:53:50 +02:00
|
|
|
{
|
2006-05-29 17:01:58 +02:00
|
|
|
thd->options |= OPTION_QUOTE_SHOW_CREATE;
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
2000-12-07 14:48:05 +01:00
|
|
|
|
2006-04-25 21:53:50 +02:00
|
|
|
/* scan sorted tmps to generate sequence of DROP */
|
2006-05-29 17:01:58 +02:00
|
|
|
for (table= thd->temporary_tables; table; table= next)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2006-05-29 17:01:58 +02:00
|
|
|
if (is_user_table(table))
|
2001-06-19 13:30:12 +02:00
|
|
|
{
|
2007-02-23 12:13:55 +01:00
|
|
|
my_thread_id save_pseudo_thread_id= thd->variables.pseudo_thread_id;
|
2006-04-25 21:53:50 +02:00
|
|
|
/* Set pseudo_thread_id to be that of the processed table */
|
|
|
|
thd->variables.pseudo_thread_id= tmpkeyval(thd, table);
|
2009-08-27 18:28:09 +02:00
|
|
|
String db;
|
2009-08-28 11:45:57 +02:00
|
|
|
db.append(table->s->db.str);
|
2009-08-27 18:28:09 +02:00
|
|
|
/* Loop forward through all tables that belong to a common database
|
|
|
|
within the sublist of common pseudo_thread_id to create single
|
|
|
|
DROP query
|
2007-02-23 12:13:55 +01:00
|
|
|
*/
|
2006-05-29 17:01:58 +02:00
|
|
|
for (s_query.length(stub_len);
|
|
|
|
table && is_user_table(table) &&
|
2009-08-27 18:28:09 +02:00
|
|
|
tmpkeyval(thd, table) == thd->variables.pseudo_thread_id &&
|
2009-08-28 11:45:57 +02:00
|
|
|
table->s->db.length == db.length() &&
|
|
|
|
strcmp(table->s->db.str, db.ptr()) == 0;
|
2006-04-25 21:53:50 +02:00
|
|
|
table= next)
|
|
|
|
{
|
2006-05-29 17:01:58 +02:00
|
|
|
/*
|
2009-08-27 18:28:09 +02:00
|
|
|
We are going to add ` around the table names and possible more
|
|
|
|
due to special characters
|
2006-05-29 17:01:58 +02:00
|
|
|
*/
|
|
|
|
append_identifier(thd, &s_query, table->s->table_name.str,
|
|
|
|
strlen(table->s->table_name.str));
|
2007-03-28 11:18:06 +02:00
|
|
|
s_query.append(',');
|
2006-04-25 21:53:50 +02:00
|
|
|
next= table->next;
|
|
|
|
close_temporary(table, 1, 1);
|
|
|
|
}
|
|
|
|
thd->clear_error();
|
2006-05-29 17:01:58 +02:00
|
|
|
CHARSET_INFO *cs_save= thd->variables.character_set_client;
|
|
|
|
thd->variables.character_set_client= system_charset_info;
|
|
|
|
Query_log_event qinfo(thd, s_query.ptr(),
|
|
|
|
s_query.length() - 1 /* to remove trailing ',' */,
|
2009-05-30 15:32:28 +02:00
|
|
|
0, FALSE, 0);
|
2009-08-27 18:28:09 +02:00
|
|
|
qinfo.db= db.ptr();
|
2009-11-03 10:00:41 +01:00
|
|
|
qinfo.db_len= db.length();
|
2006-05-29 17:01:58 +02:00
|
|
|
thd->variables.character_set_client= cs_save;
|
2006-04-25 21:53:50 +02:00
|
|
|
mysql_bin_log.write(&qinfo);
|
2007-02-23 12:13:55 +01:00
|
|
|
thd->variables.pseudo_thread_id= save_pseudo_thread_id;
|
2006-04-25 21:53:50 +02:00
|
|
|
}
|
2006-05-29 17:01:58 +02:00
|
|
|
else
|
2006-04-25 21:53:50 +02:00
|
|
|
{
|
|
|
|
next= table->next;
|
|
|
|
close_temporary(table, 1, 1);
|
2001-06-19 13:30:12 +02:00
|
|
|
}
|
2005-02-16 23:23:20 +01:00
|
|
|
}
|
2006-05-29 17:01:58 +02:00
|
|
|
if (!was_quote_show)
|
2006-11-27 00:47:38 +01:00
|
|
|
thd->options&= ~OPTION_QUOTE_SHOW_CREATE; /* restore option */
|
2000-07-31 21:29:14 +02:00
|
|
|
thd->temporary_tables=0;
|
|
|
|
}
|
|
|
|
|
2002-11-30 18:26:18 +01:00
|
|
|
/*
|
2004-09-03 20:43:04 +02:00
|
|
|
Find table in list.
|
2002-11-30 18:26:18 +01:00
|
|
|
|
|
|
|
SYNOPSIS
|
2004-09-03 20:43:04 +02:00
|
|
|
find_table_in_list()
|
2004-09-14 18:28:29 +02:00
|
|
|
table Pointer to table list
|
2004-09-03 20:43:04 +02:00
|
|
|
offset Offset to which list in table structure to use
|
2004-09-14 18:28:29 +02:00
|
|
|
db_name Data base name
|
|
|
|
table_name Table name
|
2004-07-16 00:15:55 +02:00
|
|
|
|
2004-09-03 20:43:04 +02:00
|
|
|
NOTES:
|
|
|
|
This is called by find_table_in_local_list() and
|
|
|
|
find_table_in_global_list().
|
2004-07-16 00:15:55 +02:00
|
|
|
|
|
|
|
RETURN VALUES
|
|
|
|
NULL Table not found
|
|
|
|
# Pointer to found table.
|
|
|
|
*/
|
|
|
|
|
2004-09-03 20:43:04 +02:00
|
|
|
TABLE_LIST *find_table_in_list(TABLE_LIST *table,
|
2007-07-06 14:18:49 +02:00
|
|
|
TABLE_LIST *TABLE_LIST::*link,
|
2004-09-03 20:43:04 +02:00
|
|
|
const char *db_name,
|
|
|
|
const char *table_name)
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2006-02-25 16:46:30 +01:00
|
|
|
for (; table; table= table->*link )
|
2004-09-03 20:43:04 +02:00
|
|
|
{
|
2005-03-28 14:13:31 +02:00
|
|
|
if ((table->table == 0 || table->table->s->tmp_table == NO_TMP_TABLE) &&
|
|
|
|
strcmp(table->db, db_name) == 0 &&
|
|
|
|
strcmp(table->table_name, table_name) == 0)
|
|
|
|
break;
|
2004-09-03 20:43:04 +02:00
|
|
|
}
|
2004-07-16 00:15:55 +02:00
|
|
|
return table;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2004-09-08 09:18:04 +02:00
|
|
|
/*
|
2005-04-04 23:32:48 +02:00
|
|
|
Test that table is unique (It's only exists once in the table list)
|
2004-09-08 09:18:04 +02:00
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
unique_table()
|
2005-12-22 13:48:00 +01:00
|
|
|
thd thread handle
|
2005-04-04 23:32:48 +02:00
|
|
|
table table which should be checked
|
2005-03-28 14:13:31 +02:00
|
|
|
table_list list of tables
|
2007-03-01 22:09:22 +01:00
|
|
|
check_alias whether to check tables' aliases
|
2005-03-28 14:13:31 +02:00
|
|
|
|
|
|
|
NOTE: to exclude derived tables from check we use following mechanism:
|
|
|
|
a) during derived table processing set THD::derived_tables_processing
|
|
|
|
b) JOIN::prepare set SELECT::exclude_from_table_unique_test if
|
|
|
|
THD::derived_tables_processing set. (we can't use JOIN::execute
|
|
|
|
because for PS we perform only JOIN::prepare, but we can't set this
|
|
|
|
flag in JOIN::prepare if we are not sure that we are in derived table
|
|
|
|
processing loop, because multi-update call fix_fields() for some its
|
|
|
|
items (which mean JOIN::prepare for subqueries) before unique_table
|
|
|
|
call to detect which tables should be locked for write).
|
|
|
|
c) unique_table skip all tables which belong to SELECT with
|
|
|
|
SELECT::exclude_from_table_unique_test set.
|
|
|
|
Also SELECT::exclude_from_table_unique_test used to exclude from check
|
|
|
|
tables of main SELECT of multi-delete and multi-update
|
|
|
|
|
2006-10-10 11:44:04 +02:00
|
|
|
We also skip tables with TABLE_LIST::prelocking_placeholder set,
|
|
|
|
because we want to allow SELECTs from them, and their modification
|
|
|
|
will rise the error anyway.
|
|
|
|
|
2005-03-28 14:13:31 +02:00
|
|
|
TODO: when we will have table/view change detection we can do this check
|
|
|
|
only once for PS/SP
|
2004-09-08 09:18:04 +02:00
|
|
|
|
|
|
|
RETURN
|
2004-09-10 01:22:44 +02:00
|
|
|
found duplicate
|
|
|
|
0 if table is unique
|
2004-09-08 09:18:04 +02:00
|
|
|
*/
|
|
|
|
|
2007-03-01 22:09:22 +01:00
|
|
|
TABLE_LIST* unique_table(THD *thd, TABLE_LIST *table, TABLE_LIST *table_list,
|
|
|
|
bool check_alias)
|
2004-09-08 09:18:04 +02:00
|
|
|
{
|
2004-09-10 01:22:44 +02:00
|
|
|
TABLE_LIST *res;
|
2007-03-01 22:09:22 +01:00
|
|
|
const char *d_name, *t_name, *t_alias;
|
2005-02-26 23:46:40 +01:00
|
|
|
DBUG_ENTER("unique_table");
|
|
|
|
DBUG_PRINT("enter", ("table alias: %s", table->alias));
|
2005-04-03 00:23:45 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
If this function called for query which update table (INSERT/UPDATE/...)
|
|
|
|
then we have in table->table pointer to TABLE object which we are
|
|
|
|
updating even if it is VIEW so we need TABLE_LIST of this TABLE object
|
|
|
|
to get right names (even if lower_case_table_names used).
|
|
|
|
|
|
|
|
If this function called for CREATE command that we have not opened table
|
|
|
|
(table->table equal to 0) and right names is in current TABLE_LIST
|
|
|
|
object.
|
|
|
|
*/
|
|
|
|
if (table->table)
|
2004-09-08 09:18:04 +02:00
|
|
|
{
|
2005-04-03 00:23:45 +02:00
|
|
|
/* temporary table is always unique */
|
|
|
|
if (table->table && table->table->s->tmp_table != NO_TMP_TABLE)
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
table= table->find_underlying_table(table->table);
|
|
|
|
/*
|
|
|
|
as far as we have table->table we have to find real TABLE_LIST of
|
|
|
|
it in underlying tables
|
|
|
|
*/
|
|
|
|
DBUG_ASSERT(table);
|
2004-09-08 09:18:04 +02:00
|
|
|
}
|
2005-04-03 00:23:45 +02:00
|
|
|
d_name= table->db;
|
|
|
|
t_name= table->table_name;
|
2007-03-01 22:09:22 +01:00
|
|
|
t_alias= table->alias;
|
2004-10-25 16:32:28 +02:00
|
|
|
|
|
|
|
DBUG_PRINT("info", ("real table: %s.%s", d_name, t_name));
|
2005-06-01 15:35:09 +02:00
|
|
|
for (;;)
|
2004-09-10 01:22:44 +02:00
|
|
|
{
|
2009-11-24 19:55:16 +01:00
|
|
|
/*
|
|
|
|
Table is unique if it is present only once in the global list
|
|
|
|
of tables and once in the list of table locks.
|
|
|
|
*/
|
|
|
|
if (! (res= find_table_in_global_list(table_list, d_name, t_name)) &&
|
|
|
|
! (res= mysql_lock_have_duplicate(thd, table, table_list)))
|
|
|
|
break;
|
|
|
|
|
|
|
|
/* Skip if same underlying table. */
|
|
|
|
if (res->table && (res->table == table->table))
|
|
|
|
goto next;
|
|
|
|
|
|
|
|
/* Skip if table alias does not match. */
|
|
|
|
if (check_alias)
|
|
|
|
{
|
|
|
|
if (lower_case_table_names ?
|
2007-03-02 11:25:41 +01:00
|
|
|
my_strcasecmp(files_charset_info, t_alias, res->alias) :
|
2009-11-24 19:55:16 +01:00
|
|
|
strcmp(t_alias, res->alias))
|
|
|
|
goto next;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
Skip if marked to be excluded (could be a derived table) or if
|
|
|
|
entry is a prelocking placeholder.
|
|
|
|
*/
|
|
|
|
if (res->select_lex &&
|
|
|
|
!res->select_lex->exclude_from_table_unique_test &&
|
|
|
|
!res->prelocking_placeholder)
|
2004-10-25 16:32:28 +02:00
|
|
|
break;
|
2009-11-24 19:55:16 +01:00
|
|
|
|
2005-03-28 14:13:31 +02:00
|
|
|
/*
|
2006-10-10 11:44:04 +02:00
|
|
|
If we found entry of this table or table of SELECT which already
|
2005-03-28 14:13:31 +02:00
|
|
|
processed in derived table or top select of multi-update/multi-delete
|
2006-10-10 11:44:04 +02:00
|
|
|
(exclude_from_table_unique_test) or prelocking placeholder.
|
2005-03-28 14:13:31 +02:00
|
|
|
*/
|
2009-11-24 19:55:16 +01:00
|
|
|
next:
|
2004-10-25 16:32:28 +02:00
|
|
|
table_list= res->next_global;
|
2005-03-28 14:13:31 +02:00
|
|
|
DBUG_PRINT("info",
|
|
|
|
("found same copy of table or table which we should skip"));
|
2004-09-10 01:22:44 +02:00
|
|
|
}
|
2004-10-25 16:32:28 +02:00
|
|
|
DBUG_RETURN(res);
|
2004-09-08 09:18:04 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-08-02 21:54:49 +02:00
|
|
|
/*
|
|
|
|
Issue correct error message in case we found 2 duplicate tables which
|
|
|
|
prevent some update operation
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
update_non_unique_table_error()
|
|
|
|
update table which we try to update
|
|
|
|
operation name of update operation
|
|
|
|
duplicate duplicate table which we found
|
|
|
|
|
|
|
|
NOTE:
|
|
|
|
here we hide view underlying tables if we have them
|
|
|
|
*/
|
|
|
|
|
|
|
|
void update_non_unique_table_error(TABLE_LIST *update,
|
|
|
|
const char *operation,
|
|
|
|
TABLE_LIST *duplicate)
|
|
|
|
{
|
|
|
|
update= update->top_table();
|
|
|
|
duplicate= duplicate->top_table();
|
|
|
|
if (!update->view || !duplicate->view ||
|
|
|
|
update->view == duplicate->view ||
|
|
|
|
update->view_name.length != duplicate->view_name.length ||
|
|
|
|
update->view_db.length != duplicate->view_db.length ||
|
|
|
|
my_strcasecmp(table_alias_charset,
|
|
|
|
update->view_name.str, duplicate->view_name.str) != 0 ||
|
|
|
|
my_strcasecmp(table_alias_charset,
|
|
|
|
update->view_db.str, duplicate->view_db.str) != 0)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
it is not the same view repeated (but it can be parts of the same copy
|
|
|
|
of view), so we have to hide underlying tables.
|
|
|
|
*/
|
|
|
|
if (update->view)
|
|
|
|
{
|
2006-09-28 23:00:18 +02:00
|
|
|
/* Issue the ER_NON_INSERTABLE_TABLE error for an INSERT */
|
2005-08-02 21:54:49 +02:00
|
|
|
if (update->view == duplicate->view)
|
2006-09-28 23:00:18 +02:00
|
|
|
my_error(!strncmp(operation, "INSERT", 6) ?
|
|
|
|
ER_NON_INSERTABLE_TABLE : ER_NON_UPDATABLE_TABLE, MYF(0),
|
|
|
|
update->alias, operation);
|
2005-08-02 21:54:49 +02:00
|
|
|
else
|
|
|
|
my_error(ER_VIEW_PREVENT_UPDATE, MYF(0),
|
|
|
|
(duplicate->view ? duplicate->alias : update->alias),
|
|
|
|
operation, update->alias);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
if (duplicate->view)
|
|
|
|
{
|
|
|
|
my_error(ER_VIEW_PREVENT_UPDATE, MYF(0), duplicate->alias, operation,
|
|
|
|
update->alias);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
my_error(ER_UPDATE_TABLE_USED, MYF(0), update->alias);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
TABLE *find_temporary_table(THD *thd, const char *db, const char *table_name)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
TABLE_LIST table_list;
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
table_list.db= (char*) db;
|
|
|
|
table_list.table_name= (char*) table_name;
|
|
|
|
return find_temporary_table(thd, &table_list);
|
|
|
|
}
|
2000-12-07 14:48:05 +01:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
|
|
|
|
TABLE *find_temporary_table(THD *thd, TABLE_LIST *table_list)
|
|
|
|
{
|
|
|
|
char key[MAX_DBKEY_LENGTH];
|
|
|
|
uint key_length;
|
|
|
|
TABLE *table;
|
2006-08-02 17:57:06 +02:00
|
|
|
DBUG_ENTER("find_temporary_table");
|
|
|
|
DBUG_PRINT("enter", ("table: '%s'.'%s'",
|
|
|
|
table_list->db, table_list->table_name));
|
2005-11-23 21:45:02 +01:00
|
|
|
|
|
|
|
key_length= create_table_def_key(thd, key, table_list, 1);
|
|
|
|
for (table=thd->temporary_tables ; table ; table= table->next)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
if (table->s->table_cache_key.length == key_length &&
|
|
|
|
!memcmp(table->s->table_cache_key.str, key, key_length))
|
2007-02-23 12:13:55 +01:00
|
|
|
{
|
|
|
|
DBUG_PRINT("info",
|
|
|
|
("Found table. server_id: %u pseudo_thread_id: %lu",
|
|
|
|
(uint) thd->server_id,
|
|
|
|
(ulong) thd->variables.pseudo_thread_id));
|
2006-08-02 17:57:06 +02:00
|
|
|
DBUG_RETURN(table);
|
2007-02-23 12:13:55 +01:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2007-02-23 12:13:55 +01:00
|
|
|
DBUG_RETURN(0); // Not a temporary table
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
|
A fix and a test case for Bug#24918 drop table and lock / inconsistent
between perm and temp tables. Review fixes.
The original bug report complains that if we locked a temporary table
with LOCK TABLES statement, we would not leave LOCK TABLES mode
when this temporary table is dropped.
Additionally, the bug was escalated when it was discovered than
when a temporary transactional table that was previously
locked with LOCK TABLES statement was dropped, futher actions with
this table, such as UNLOCK TABLES, would lead to a crash.
The problem originates from incomplete support of transactional temporary
tables. When we added calls to handler::store_lock()/handler::external_lock()
to operations that work with such tables, we only covered the normal
server code flow and did not cover LOCK TABLES mode.
In LOCK TABLES mode, ::external_lock(LOCK) would sometimes be called without
matching ::external_lock(UNLOCK), e.g. when a transactional temporary table
was dropped. Additionally, this table would be left in the list of LOCKed
TABLES.
The patch aims to address this inadequacy. Now, whenever an instance
of 'handler' is destroyed, we assert that it was priorly
external_lock(UNLOCK)-ed. All the places that violate this assert
were fixed.
This patch introduces no changes in behavior -- the discrepancy in
behavior will be fixed when we start calling ::store_lock()/::external_lock()
for all tables, regardless whether they are transactional or not,
temporary or not.
2007-07-27 14:37:29 +02:00
|
|
|
/**
|
|
|
|
Drop a temporary table.
|
|
|
|
|
|
|
|
Try to locate the table in the list of thd->temporary_tables.
|
|
|
|
If the table is found:
|
2007-11-01 21:52:56 +01:00
|
|
|
- if the table is being used by some outer statement, fail.
|
2009-12-01 15:39:03 +01:00
|
|
|
- if the table is locked with LOCK TABLES or by prelocking,
|
|
|
|
unlock it and remove it from the list of locked tables
|
|
|
|
(THD::lock). Currently only transactional temporary tables
|
|
|
|
are locked.
|
A fix and a test case for Bug#24918 drop table and lock / inconsistent
between perm and temp tables. Review fixes.
The original bug report complains that if we locked a temporary table
with LOCK TABLES statement, we would not leave LOCK TABLES mode
when this temporary table is dropped.
Additionally, the bug was escalated when it was discovered than
when a temporary transactional table that was previously
locked with LOCK TABLES statement was dropped, futher actions with
this table, such as UNLOCK TABLES, would lead to a crash.
The problem originates from incomplete support of transactional temporary
tables. When we added calls to handler::store_lock()/handler::external_lock()
to operations that work with such tables, we only covered the normal
server code flow and did not cover LOCK TABLES mode.
In LOCK TABLES mode, ::external_lock(LOCK) would sometimes be called without
matching ::external_lock(UNLOCK), e.g. when a transactional temporary table
was dropped. Additionally, this table would be left in the list of LOCKed
TABLES.
The patch aims to address this inadequacy. Now, whenever an instance
of 'handler' is destroyed, we assert that it was priorly
external_lock(UNLOCK)-ed. All the places that violate this assert
were fixed.
This patch introduces no changes in behavior -- the discrepancy in
behavior will be fixed when we start calling ::store_lock()/::external_lock()
for all tables, regardless whether they are transactional or not,
temporary or not.
2007-07-27 14:37:29 +02:00
|
|
|
- Close the temporary table, remove its .FRM
|
|
|
|
- remove the table from the list of temporary tables
|
|
|
|
|
|
|
|
This function is used to drop user temporary tables, as well as
|
|
|
|
internal tables created in CREATE TEMPORARY TABLE ... SELECT
|
|
|
|
or ALTER TABLE. Even though part of the work done by this function
|
|
|
|
is redundant when the table is internal, as long as we
|
|
|
|
link both internal and user temporary tables into the same
|
|
|
|
thd->temporary_tables list, it's impossible to tell here whether
|
|
|
|
we're dealing with an internal or a user temporary table.
|
|
|
|
|
2007-11-01 21:52:56 +01:00
|
|
|
@retval 0 the table was found and dropped successfully.
|
|
|
|
@retval 1 the table was not found in the list of temporary tables
|
|
|
|
of this thread
|
|
|
|
@retval -1 the table is in use by a outer query
|
2005-11-23 21:45:02 +01:00
|
|
|
*/
|
|
|
|
|
2007-11-01 21:52:56 +01:00
|
|
|
int drop_temporary_table(THD *thd, TABLE_LIST *table_list)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
TABLE *table;
|
2007-11-01 21:52:56 +01:00
|
|
|
DBUG_ENTER("drop_temporary_table");
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
DBUG_PRINT("tmptable", ("closing table: '%s'.'%s'",
|
|
|
|
table_list->db, table_list->table_name));
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
if (!(table= find_temporary_table(thd, table_list)))
|
2007-11-01 21:52:56 +01:00
|
|
|
DBUG_RETURN(1);
|
|
|
|
|
|
|
|
/* Table might be in use by some outer statement. */
|
|
|
|
if (table->query_id && table->query_id != thd->query_id)
|
|
|
|
{
|
|
|
|
my_error(ER_CANT_REOPEN_TABLE, MYF(0), table->alias);
|
|
|
|
DBUG_RETURN(-1);
|
|
|
|
}
|
|
|
|
|
A fix and a test case for Bug#24918 drop table and lock / inconsistent
between perm and temp tables. Review fixes.
The original bug report complains that if we locked a temporary table
with LOCK TABLES statement, we would not leave LOCK TABLES mode
when this temporary table is dropped.
Additionally, the bug was escalated when it was discovered than
when a temporary transactional table that was previously
locked with LOCK TABLES statement was dropped, futher actions with
this table, such as UNLOCK TABLES, would lead to a crash.
The problem originates from incomplete support of transactional temporary
tables. When we added calls to handler::store_lock()/handler::external_lock()
to operations that work with such tables, we only covered the normal
server code flow and did not cover LOCK TABLES mode.
In LOCK TABLES mode, ::external_lock(LOCK) would sometimes be called without
matching ::external_lock(UNLOCK), e.g. when a transactional temporary table
was dropped. Additionally, this table would be left in the list of LOCKed
TABLES.
The patch aims to address this inadequacy. Now, whenever an instance
of 'handler' is destroyed, we assert that it was priorly
external_lock(UNLOCK)-ed. All the places that violate this assert
were fixed.
This patch introduces no changes in behavior -- the discrepancy in
behavior will be fixed when we start calling ::store_lock()/::external_lock()
for all tables, regardless whether they are transactional or not,
temporary or not.
2007-07-27 14:37:29 +02:00
|
|
|
/*
|
|
|
|
If LOCK TABLES list is not empty and contains this table,
|
|
|
|
unlock the table and remove the table from this list.
|
|
|
|
*/
|
2009-12-02 16:22:15 +01:00
|
|
|
mysql_lock_remove(thd, thd->lock, table);
|
2005-11-23 21:45:02 +01:00
|
|
|
close_temporary_table(thd, table, 1, 1);
|
2007-11-01 21:52:56 +01:00
|
|
|
DBUG_RETURN(0);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
/*
|
2006-06-19 19:14:14 +02:00
|
|
|
unlink from thd->temporary tables and close temporary table
|
2005-11-23 21:45:02 +01:00
|
|
|
*/
|
|
|
|
|
|
|
|
void close_temporary_table(THD *thd, TABLE *table,
|
|
|
|
bool free_share, bool delete_table)
|
|
|
|
{
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
DBUG_ENTER("close_temporary_table");
|
|
|
|
DBUG_PRINT("tmptable", ("closing table: '%s'.'%s' 0x%lx alias: '%s'",
|
|
|
|
table->s->db.str, table->s->table_name.str,
|
|
|
|
(long) table, table->alias));
|
|
|
|
|
2006-06-19 19:14:14 +02:00
|
|
|
if (table->prev)
|
|
|
|
{
|
|
|
|
table->prev->next= table->next;
|
|
|
|
if (table->prev->next)
|
|
|
|
table->next->prev= table->prev;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/* removing the item from the list */
|
|
|
|
DBUG_ASSERT(table == thd->temporary_tables);
|
|
|
|
/*
|
|
|
|
slave must reset its temporary list pointer to zero to exclude
|
|
|
|
passing non-zero value to end_slave via rli->save_temporary_tables
|
|
|
|
when no temp tables opened, see an invariant below.
|
|
|
|
*/
|
|
|
|
thd->temporary_tables= table->next;
|
|
|
|
if (thd->temporary_tables)
|
|
|
|
table->next->prev= 0;
|
|
|
|
}
|
2005-11-23 21:45:02 +01:00
|
|
|
if (thd->slave_thread)
|
2006-06-19 19:14:14 +02:00
|
|
|
{
|
|
|
|
/* natural invariant of temporary_tables */
|
|
|
|
DBUG_ASSERT(slave_open_temp_tables || !thd->temporary_tables);
|
2005-11-23 21:45:02 +01:00
|
|
|
slave_open_temp_tables--;
|
2006-06-19 19:14:14 +02:00
|
|
|
}
|
2005-11-23 21:45:02 +01:00
|
|
|
close_temporary(table, free_share, delete_table);
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
DBUG_VOID_RETURN;
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Close and delete a temporary table
|
|
|
|
|
|
|
|
NOTE
|
|
|
|
This dosn't unlink table from thd->temporary
|
|
|
|
If this is needed, use close_temporary_table()
|
|
|
|
*/
|
|
|
|
|
|
|
|
void close_temporary(TABLE *table, bool free_share, bool delete_table)
|
|
|
|
{
|
2007-03-02 17:43:45 +01:00
|
|
|
handlerton *table_type= table->s->db_type();
|
2005-11-23 21:45:02 +01:00
|
|
|
DBUG_ENTER("close_temporary");
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
DBUG_PRINT("tmptable", ("closing table: '%s'.'%s'",
|
|
|
|
table->s->db.str, table->s->table_name.str));
|
2005-11-23 21:45:02 +01:00
|
|
|
|
|
|
|
free_io_cache(table);
|
|
|
|
closefrm(table, 0);
|
|
|
|
if (delete_table)
|
|
|
|
rm_temporary_table(table_type, table->s->path.str);
|
|
|
|
if (free_share)
|
|
|
|
{
|
|
|
|
free_table_share(table->s);
|
|
|
|
my_free((char*) table,MYF(0));
|
|
|
|
}
|
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2003-10-29 14:23:35 +01:00
|
|
|
/*
|
|
|
|
Used by ALTER TABLE when the table is a temporary one. It changes something
|
|
|
|
only if the ALTER contained a RENAME clause (otherwise, table_name is the old
|
|
|
|
name).
|
|
|
|
Prepares a table cache key, which is the concatenation of db, table_name and
|
|
|
|
thd->slave_proxy_id, separated by '\0'.
|
|
|
|
*/
|
2005-01-06 12:00:13 +01:00
|
|
|
|
2000-11-26 07:30:02 +01:00
|
|
|
bool rename_temporary_table(THD* thd, TABLE *table, const char *db,
|
2000-07-31 21:29:14 +02:00
|
|
|
const char *table_name)
|
|
|
|
{
|
|
|
|
char *key;
|
2006-08-21 17:02:11 +02:00
|
|
|
uint key_length;
|
2005-01-06 12:00:13 +01:00
|
|
|
TABLE_SHARE *share= table->s;
|
2005-11-23 21:45:02 +01:00
|
|
|
TABLE_LIST table_list;
|
|
|
|
DBUG_ENTER("rename_temporary_table");
|
|
|
|
|
2006-08-21 17:02:11 +02:00
|
|
|
if (!(key=(char*) alloc_root(&share->mem_root, MAX_DBKEY_LENGTH)))
|
2005-11-23 21:45:02 +01:00
|
|
|
DBUG_RETURN(1); /* purecov: inspected */
|
|
|
|
|
|
|
|
table_list.db= (char*) db;
|
|
|
|
table_list.table_name= (char*) table_name;
|
2006-08-21 17:02:11 +02:00
|
|
|
key_length= create_table_def_key(thd, key, &table_list, 1);
|
|
|
|
share->set_table_cache_key(key, key_length);
|
2005-11-23 21:45:02 +01:00
|
|
|
DBUG_RETURN(0);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-01 14:27:03 +01:00
|
|
|
/**
|
|
|
|
Force all other threads to stop using the table by upgrading
|
|
|
|
metadata lock on it and remove unused TABLE instances from cache.
|
|
|
|
|
|
|
|
@param thd Thread handler
|
|
|
|
@param table Table to remove from cache
|
|
|
|
@param function HA_EXTRA_PREPARE_FOR_DROP if table is to be deleted
|
|
|
|
HA_EXTRA_FORCE_REOPEN if table is not be used
|
|
|
|
HA_EXTRA_PREPARE_FOR_RENAME if table is to be renamed
|
|
|
|
|
|
|
|
@note When returning, the table will be unusable for other threads
|
|
|
|
until metadata lock is downgraded.
|
|
|
|
|
|
|
|
@retval FALSE Success.
|
|
|
|
@retval TRUE Failure (e.g. because thread was killed).
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool wait_while_table_is_used(THD *thd, TABLE *table,
|
|
|
|
enum ha_extra_function function)
|
|
|
|
{
|
|
|
|
enum thr_lock_type old_lock_type;
|
|
|
|
|
|
|
|
DBUG_ENTER("wait_while_table_is_used");
|
|
|
|
DBUG_PRINT("enter", ("table: '%s' share: 0x%lx db_stat: %u version: %lu",
|
|
|
|
table->s->table_name.str, (ulong) table->s,
|
|
|
|
table->db_stat, table->s->version));
|
|
|
|
|
|
|
|
(void) table->file->extra(function);
|
|
|
|
|
|
|
|
old_lock_type= table->reginfo.lock_type;
|
|
|
|
mysql_lock_abort(thd, table, TRUE); /* end threads waiting on lock */
|
|
|
|
|
2009-12-04 00:52:05 +01:00
|
|
|
if (table->mdl_ticket->upgrade_shared_lock_to_exclusive())
|
2009-12-01 14:27:03 +01:00
|
|
|
{
|
|
|
|
mysql_lock_downgrade_write(thd, table, old_lock_type);
|
|
|
|
DBUG_RETURN(TRUE);
|
|
|
|
}
|
|
|
|
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
2009-12-01 20:13:01 +01:00
|
|
|
tdc_remove_table(thd, TDC_RT_REMOVE_NOT_OWN,
|
|
|
|
table->s->db.str, table->s->table_name.str);
|
2009-12-01 14:27:03 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
DBUG_RETURN(FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
2009-12-02 16:22:15 +01:00
|
|
|
Close a and drop a just created table in CREATE TABLE ... SELECT.
|
|
|
|
|
|
|
|
@param thd Thread handle
|
|
|
|
@param table TABLE object for the table to be dropped
|
|
|
|
@param db_name Name of database for this table
|
|
|
|
@param table_name Name of this table
|
|
|
|
|
|
|
|
This routine assumes that the table to be closed is open only
|
|
|
|
by the calling thread, so we needn't wait until other threads
|
|
|
|
close the table. It also assumes that the table is first
|
|
|
|
in thd->open_ables and a data lock on it, if any, has been
|
|
|
|
released. To sum up, it's tuned to work with
|
|
|
|
CREATE TABLE ... SELECT and CREATE TABLE .. SELECT only.
|
|
|
|
Note, that currently CREATE TABLE ... SELECT is not supported
|
|
|
|
under LOCK TABLES. This function, still, can be called in
|
|
|
|
prelocked mode, e.g. if we do CREATE TABLE .. SELECT f1();
|
2007-05-11 19:51:03 +02:00
|
|
|
*/
|
|
|
|
|
|
|
|
void drop_open_table(THD *thd, TABLE *table, const char *db_name,
|
|
|
|
const char *table_name)
|
|
|
|
{
|
|
|
|
if (table->s->tmp_table)
|
|
|
|
close_temporary_table(thd, table, 1, 1);
|
|
|
|
else
|
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
DBUG_ASSERT(table == thd->open_tables);
|
|
|
|
|
2007-05-15 15:58:26 +02:00
|
|
|
handlerton *table_type= table->s->db_type();
|
2009-12-02 16:22:15 +01:00
|
|
|
/* Ensure the table is removed from the cache. */
|
|
|
|
table->s->version= 0;
|
|
|
|
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_lock(&LOCK_open);
|
2009-12-02 16:22:15 +01:00
|
|
|
close_thread_table(thd, &thd->open_tables);
|
2007-05-11 19:51:03 +02:00
|
|
|
quick_rm_table(table_type, db_name, table_name, 0);
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2007-05-11 19:51:03 +02:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2000-08-28 15:43:58 +02:00
|
|
|
/*
|
2005-11-23 21:45:02 +01:00
|
|
|
Wait for condition but allow the user to send a kill to mysqld
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
wait_for_condition()
|
|
|
|
thd Thread handler
|
|
|
|
mutex mutex that is currently hold that is associated with condition
|
|
|
|
Will be unlocked on return
|
|
|
|
cond Condition to wait for
|
2000-07-31 21:29:14 +02:00
|
|
|
*/
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
void wait_for_condition(THD *thd, pthread_mutex_t *mutex, pthread_cond_t *cond)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
/* Wait until the current table is up to date */
|
|
|
|
const char *proc_info;
|
2005-11-23 21:45:02 +01:00
|
|
|
thd->mysys_var->current_mutex= mutex;
|
|
|
|
thd->mysys_var->current_cond= cond;
|
2000-07-31 21:29:14 +02:00
|
|
|
proc_info=thd->proc_info;
|
2007-02-22 16:03:08 +01:00
|
|
|
thd_proc_info(thd, "Waiting for table");
|
2006-05-30 15:07:49 +02:00
|
|
|
DBUG_ENTER("wait_for_condition");
|
2001-09-01 09:38:16 +02:00
|
|
|
if (!thd->killed)
|
2005-11-23 21:45:02 +01:00
|
|
|
(void) pthread_cond_wait(cond, mutex);
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
/*
|
|
|
|
We must unlock mutex first to avoid deadlock becasue conditions are
|
|
|
|
sent to this thread by doing locks in the following order:
|
|
|
|
lock(mysys_var->mutex)
|
|
|
|
lock(mysys_var->current_mutex)
|
|
|
|
|
|
|
|
One by effect of this that one can only use wait_for_condition with
|
|
|
|
condition variables that are guranteed to not disapper (freed) even if this
|
|
|
|
mutex is unlocked
|
|
|
|
*/
|
|
|
|
|
|
|
|
pthread_mutex_unlock(mutex);
|
2000-07-31 21:29:14 +02:00
|
|
|
pthread_mutex_lock(&thd->mysys_var->mutex);
|
|
|
|
thd->mysys_var->current_mutex= 0;
|
|
|
|
thd->mysys_var->current_cond= 0;
|
2007-02-22 16:03:08 +01:00
|
|
|
thd_proc_info(thd, proc_info);
|
2000-07-31 21:29:14 +02:00
|
|
|
pthread_mutex_unlock(&thd->mysys_var->mutex);
|
2006-05-24 16:21:35 +02:00
|
|
|
DBUG_VOID_RETURN;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2002-06-11 10:20:31 +02:00
|
|
|
|
2007-05-11 19:51:03 +02:00
|
|
|
/**
|
2007-10-16 21:37:31 +02:00
|
|
|
Check that table exists in table definition cache, on disk
|
|
|
|
or in some storage engine.
|
2007-05-11 19:51:03 +02:00
|
|
|
|
2007-08-15 15:43:08 +02:00
|
|
|
@param thd Thread context
|
|
|
|
@param table Table list element
|
|
|
|
@param[out] exists Out parameter which is set to TRUE if table
|
2007-05-11 19:51:03 +02:00
|
|
|
exists and to FALSE otherwise.
|
|
|
|
|
|
|
|
@note This function assumes that caller owns LOCK_open mutex.
|
2009-12-01 14:51:50 +01:00
|
|
|
It also assumes that the fact that there are no exclusive
|
|
|
|
metadata locks on the table was checked beforehand.
|
2007-05-11 19:51:03 +02:00
|
|
|
|
|
|
|
@note If there is no .FRM file for the table but it exists in one
|
|
|
|
of engines (e.g. it was created on another node of NDB cluster)
|
|
|
|
this function will fetch and create proper .FRM file for it.
|
|
|
|
|
|
|
|
@retval TRUE Some error occured
|
|
|
|
@retval FALSE No error. 'exists' out parameter set accordingly.
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool check_if_table_exists(THD *thd, TABLE_LIST *table, bool *exists)
|
|
|
|
{
|
2009-06-19 10:24:43 +02:00
|
|
|
char path[FN_REFLEN + 1];
|
2007-05-11 19:51:03 +02:00
|
|
|
int rc;
|
|
|
|
DBUG_ENTER("check_if_table_exists");
|
|
|
|
|
|
|
|
safe_mutex_assert_owner(&LOCK_open);
|
|
|
|
|
|
|
|
*exists= TRUE;
|
|
|
|
|
|
|
|
if (get_cached_table_share(table->db, table->table_name))
|
|
|
|
DBUG_RETURN(FALSE);
|
|
|
|
|
|
|
|
build_table_filename(path, sizeof(path) - 1, table->db, table->table_name,
|
|
|
|
reg_ext, 0);
|
|
|
|
|
|
|
|
if (!access(path, F_OK))
|
|
|
|
DBUG_RETURN(FALSE);
|
|
|
|
|
|
|
|
/* .FRM file doesn't exist. Check if some engine can provide it. */
|
|
|
|
|
|
|
|
rc= ha_create_table_from_engine(thd, table->db, table->table_name);
|
|
|
|
|
|
|
|
if (rc < 0)
|
|
|
|
{
|
|
|
|
/* Table does not exists in engines as well. */
|
|
|
|
*exists= FALSE;
|
|
|
|
DBUG_RETURN(FALSE);
|
|
|
|
}
|
|
|
|
else if (!rc)
|
|
|
|
{
|
|
|
|
/* Table exists in some engine and .FRM for it was created. */
|
|
|
|
DBUG_RETURN(FALSE);
|
|
|
|
}
|
|
|
|
else /* (rc > 0) */
|
|
|
|
{
|
|
|
|
my_printf_error(ER_UNKNOWN_ERROR, "Failed to open '%-.64s', error while "
|
|
|
|
"unpacking from engine", MYF(0), table->table_name);
|
|
|
|
DBUG_RETURN(TRUE);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/**
|
|
|
|
@brief Helper function used by MDL subsystem for releasing TABLE_SHARE
|
|
|
|
objects in cases when it no longer wants to cache reference to it.
|
|
|
|
*/
|
|
|
|
|
|
|
|
void table_share_release_hook(void *share)
|
|
|
|
{
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
2009-11-30 20:11:32 +01:00
|
|
|
release_table_share((TABLE_SHARE*) share);
|
2009-11-30 16:55:03 +01:00
|
|
|
broadcast_refresh();
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-02 17:15:40 +01:00
|
|
|
/**
|
|
|
|
A helper function that acquires an MDL lock for a table
|
|
|
|
being opened.
|
2009-12-02 16:22:15 +01:00
|
|
|
*/
|
|
|
|
|
|
|
|
static bool
|
|
|
|
open_table_get_mdl_lock(THD *thd, TABLE_LIST *table_list,
|
2009-12-04 00:52:05 +01:00
|
|
|
MDL_request *mdl_request,
|
2009-12-08 10:57:07 +01:00
|
|
|
Open_table_context *ot_ctx,
|
|
|
|
uint flags)
|
2009-12-02 16:22:15 +01:00
|
|
|
{
|
2009-12-08 10:57:07 +01:00
|
|
|
ot_ctx->add_request(mdl_request);
|
2009-12-02 16:22:15 +01:00
|
|
|
|
Backport of revno ## 2617.31.1, 2617.31.3, 2617.31.4, 2617.31.5,
2617.31.12, 2617.31.15, 2617.31.15, 2617.31.16, 2617.43.1
- initial changeset that introduced the fix for
Bug#989 and follow up fixes for all test suite failures
introduced in the initial changeset.
------------------------------------------------------------
revno: 2617.31.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Fri 2009-03-06 19:17:00 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
Currently the MySQL server does not keep metadata locks on
schema objects for the duration of a transaction, thus failing
to guarantee the integrity of the schema objects being used
during the transaction and to protect then from concurrent
DDL operations. This also poses a problem for replication as
a DDL operation might be replicated even thought there are
active transactions using the object being modified.
The solution is to defer the release of metadata locks until
a active transaction is either committed or rolled back. This
prevents other statements from modifying the table for the
entire duration of the transaction. This provides commitment
ordering for guaranteeing serializability across multiple
transactions.
- Incompatible change:
If MySQL's metadata locking system encounters a lock conflict,
the usual schema is to use the try and back-off technique to
avoid deadlocks -- this schema consists in releasing all locks
and trying to acquire them all in one go.
But in a transactional context this algorithm can't be utilized
as its not possible to release locks acquired during the course
of the transaction without breaking the transaction commitments.
To avoid deadlocks in this case, the ER_LOCK_DEADLOCK will be
returned if a lock conflict is encountered during a transaction.
Let's consider an example:
A transaction has two statements that modify table t1, then table
t2, and then commits. The first statement of the transaction will
acquire a shared metadata lock on table t1, and it will be kept
utill COMMIT to ensure serializability.
At the moment when the second statement attempts to acquire a
shared metadata lock on t2, a concurrent ALTER or DROP statement
might have locked t2 exclusively. The prescription of the current
locking protocol is that the acquirer of the shared lock backs off
-- gives up all his current locks and retries. This implies that
the entire multi-statement transaction has to be rolled back.
- Incompatible change:
FLUSH commands such as FLUSH PRIVILEGES and FLUSH TABLES WITH READ
LOCK won't cause locked tables to be implicitly unlocked anymore.
2009-12-05 00:02:48 +01:00
|
|
|
if (table_list->lock_strategy)
|
2009-12-02 16:22:15 +01:00
|
|
|
{
|
|
|
|
/*
|
|
|
|
In case of CREATE TABLE .. If NOT EXISTS .. SELECT, the table
|
|
|
|
may not yet exist. Let's acquire an exclusive lock for that
|
|
|
|
case. If later it turns out the table existsed, we will
|
|
|
|
downgrade the lock to shared. Note that, according to the
|
|
|
|
locking protocol, all exclusive locks must be acquired before
|
|
|
|
shared locks. This invariant is preserved here and is also
|
|
|
|
enforced by asserts in metadata locking subsystem.
|
|
|
|
*/
|
2009-12-04 00:52:05 +01:00
|
|
|
mdl_request->set_type(MDL_EXCLUSIVE);
|
2009-12-08 10:57:07 +01:00
|
|
|
DBUG_ASSERT(! thd->mdl_context.has_locks());
|
|
|
|
|
|
|
|
if (thd->mdl_context.acquire_exclusive_lock(mdl_request))
|
2009-12-02 16:22:15 +01:00
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
There is no MDL_SHARED_UPGRADABLE_HIGH_PRIO type of metadata lock so we
|
|
|
|
want to be sure that caller doesn't pass us both flags simultaneously.
|
|
|
|
*/
|
|
|
|
DBUG_ASSERT(!(flags & MYSQL_OPEN_TAKE_UPGRADABLE_MDL) ||
|
|
|
|
!(flags & MYSQL_LOCK_IGNORE_FLUSH));
|
|
|
|
|
|
|
|
if (flags & MYSQL_OPEN_TAKE_UPGRADABLE_MDL &&
|
|
|
|
table_list->lock_type >= TL_WRITE_ALLOW_WRITE)
|
2009-12-04 00:52:05 +01:00
|
|
|
mdl_request->set_type(MDL_SHARED_UPGRADABLE);
|
2009-12-02 16:22:15 +01:00
|
|
|
if (flags & MYSQL_LOCK_IGNORE_FLUSH)
|
2009-12-04 00:52:05 +01:00
|
|
|
mdl_request->set_type(MDL_SHARED_HIGH_PRIO);
|
2009-12-02 16:22:15 +01:00
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
if (thd->mdl_context.try_acquire_shared_lock(mdl_request))
|
|
|
|
return 1;
|
|
|
|
if (mdl_request->ticket == NULL)
|
2009-12-02 16:22:15 +01:00
|
|
|
{
|
2009-12-09 16:48:42 +01:00
|
|
|
if (flags & MYSQL_OPEN_FAIL_ON_MDL_CONFLICT)
|
|
|
|
my_error(ER_WARN_I_S_SKIPPED_TABLE, MYF(0), table_list->db, table_list->table_name);
|
|
|
|
else
|
|
|
|
(void) ot_ctx->request_backoff_action(Open_table_context::OT_WAIT);
|
2009-12-02 16:22:15 +01:00
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-07-13 11:48:13 +02:00
|
|
|
/*
|
|
|
|
Open a table.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
open_table()
|
2005-09-13 14:36:43 +02:00
|
|
|
thd Thread context.
|
|
|
|
table_list Open first table in list.
|
2009-11-30 16:55:03 +01:00
|
|
|
action INOUT Pointer to variable of enum_open_table_action type
|
|
|
|
which will be set according to action which is
|
|
|
|
required to remedy problem appeared during attempt
|
|
|
|
to open table.
|
2005-09-13 14:36:43 +02:00
|
|
|
flags Bitmap of flags to modify how open works:
|
|
|
|
MYSQL_LOCK_IGNORE_FLUSH - Open table even if
|
2009-11-30 16:55:03 +01:00
|
|
|
someone has done a flush or there is a pending
|
|
|
|
exclusive metadata lock requests against it
|
|
|
|
(i.e. request high priority metadata lock).
|
2005-09-13 14:36:43 +02:00
|
|
|
No version number checking is done.
|
2007-05-11 19:51:03 +02:00
|
|
|
MYSQL_OPEN_TEMPORARY_ONLY - Open only temporary
|
|
|
|
table not the base table or view.
|
2009-12-01 14:38:00 +01:00
|
|
|
MYSQL_OPEN_TAKE_UPGRADABLE_MDL - Obtain upgradable
|
|
|
|
metadata lock for tables on which we are going to
|
|
|
|
take some kind of write table-level lock.
|
2005-07-13 11:48:13 +02:00
|
|
|
|
|
|
|
IMPLEMENTATION
|
|
|
|
Uses a cache of open tables to find a table not in use.
|
|
|
|
|
Backport of revno ## 2617.31.1, 2617.31.3, 2617.31.4, 2617.31.5,
2617.31.12, 2617.31.15, 2617.31.15, 2617.31.16, 2617.43.1
- initial changeset that introduced the fix for
Bug#989 and follow up fixes for all test suite failures
introduced in the initial changeset.
------------------------------------------------------------
revno: 2617.31.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Fri 2009-03-06 19:17:00 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
Currently the MySQL server does not keep metadata locks on
schema objects for the duration of a transaction, thus failing
to guarantee the integrity of the schema objects being used
during the transaction and to protect then from concurrent
DDL operations. This also poses a problem for replication as
a DDL operation might be replicated even thought there are
active transactions using the object being modified.
The solution is to defer the release of metadata locks until
a active transaction is either committed or rolled back. This
prevents other statements from modifying the table for the
entire duration of the transaction. This provides commitment
ordering for guaranteeing serializability across multiple
transactions.
- Incompatible change:
If MySQL's metadata locking system encounters a lock conflict,
the usual schema is to use the try and back-off technique to
avoid deadlocks -- this schema consists in releasing all locks
and trying to acquire them all in one go.
But in a transactional context this algorithm can't be utilized
as its not possible to release locks acquired during the course
of the transaction without breaking the transaction commitments.
To avoid deadlocks in this case, the ER_LOCK_DEADLOCK will be
returned if a lock conflict is encountered during a transaction.
Let's consider an example:
A transaction has two statements that modify table t1, then table
t2, and then commits. The first statement of the transaction will
acquire a shared metadata lock on table t1, and it will be kept
utill COMMIT to ensure serializability.
At the moment when the second statement attempts to acquire a
shared metadata lock on t2, a concurrent ALTER or DROP statement
might have locked t2 exclusively. The prescription of the current
locking protocol is that the acquirer of the shared lock backs off
-- gives up all his current locks and retries. This implies that
the entire multi-statement transaction has to be rolled back.
- Incompatible change:
FLUSH commands such as FLUSH PRIVILEGES and FLUSH TABLES WITH READ
LOCK won't cause locked tables to be implicitly unlocked anymore.
2009-12-05 00:02:48 +01:00
|
|
|
If TABLE_LIST::open_strategy is set to OPEN_IF_EXISTS, the table is opened
|
|
|
|
only if it exists. If the open strategy is OPEN_STUB, the underlying table
|
|
|
|
is never opened. In both cases, metadata locks are always taken according
|
|
|
|
to the lock strategy.
|
|
|
|
|
|
|
|
This function will take a exclusive metadata lock on the table if
|
|
|
|
TABLE_LIST::lock_strategy is EXCLUSIVE_DOWNGRADABLE_MDL or EXCLUSIVE_MDL.
|
|
|
|
If the lock strategy is EXCLUSIVE_DOWNGRADABLE_MDL and opening the table
|
|
|
|
is successful, the exclusive metadata lock is downgraded to a shared
|
|
|
|
lock.
|
2007-05-11 19:51:03 +02:00
|
|
|
|
2005-07-13 11:48:13 +02:00
|
|
|
RETURN
|
2009-12-01 15:58:31 +01:00
|
|
|
TRUE Open failed. "action" parameter may contain type of action
|
|
|
|
needed to remedy problem before retrying again.
|
|
|
|
FALSE Success. Members of TABLE_LIST structure are filled properly (e.g.
|
|
|
|
TABLE_LIST::table is set for real tables and TABLE_LIST::view is
|
|
|
|
set for views).
|
2005-07-13 11:48:13 +02:00
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
|
2009-12-01 15:58:31 +01:00
|
|
|
bool open_table(THD *thd, TABLE_LIST *table_list, MEM_ROOT *mem_root,
|
2009-12-08 10:57:07 +01:00
|
|
|
Open_table_context *ot_ctx, uint flags)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
reg1 TABLE *table;
|
|
|
|
char key[MAX_DBKEY_LENGTH];
|
|
|
|
uint key_length;
|
2004-07-16 00:15:55 +02:00
|
|
|
char *alias= table_list->alias;
|
2009-12-04 00:52:05 +01:00
|
|
|
MDL_request *mdl_request;
|
|
|
|
MDL_ticket *mdl_ticket;
|
2009-11-30 16:55:03 +01:00
|
|
|
int error;
|
|
|
|
TABLE_SHARE *share;
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("open_table");
|
|
|
|
|
2005-11-23 00:11:19 +01:00
|
|
|
/* an open table operation needs a lot of the stack space */
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
if (check_stack_overrun(thd, STACK_MIN_SIZE_FOR_OPEN, (uchar *)&alias))
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2005-11-23 00:11:19 +01:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
if (thd->killed)
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2005-11-23 21:45:02 +01:00
|
|
|
|
|
|
|
key_length= (create_table_def_key(thd, key, table_list, 1) -
|
|
|
|
TMP_TABLE_KEY_EXTRA);
|
2000-12-07 14:48:05 +01:00
|
|
|
|
2007-02-28 14:03:38 +01:00
|
|
|
/*
|
|
|
|
Unless requested otherwise, try to resolve this table in the list
|
|
|
|
of temporary tables of this thread. In MySQL temporary tables
|
|
|
|
are always thread-local and "shadow" possible base tables with the
|
|
|
|
same name. This block implements the behaviour.
|
|
|
|
TODO: move this block into a separate function.
|
|
|
|
*/
|
2009-12-02 16:22:15 +01:00
|
|
|
if (!table_list->skip_temporary && ! (flags & MYSQL_OPEN_SKIP_TEMPORARY))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2004-07-16 00:15:55 +02:00
|
|
|
for (table= thd->temporary_tables; table ; table=table->next)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
if (table->s->table_cache_key.length == key_length +
|
|
|
|
TMP_TABLE_KEY_EXTRA &&
|
|
|
|
!memcmp(table->s->table_cache_key.str, key,
|
2004-10-29 18:26:52 +02:00
|
|
|
key_length + TMP_TABLE_KEY_EXTRA))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2007-02-28 14:03:38 +01:00
|
|
|
/*
|
|
|
|
We're trying to use the same temporary table twice in a query.
|
|
|
|
Right now we don't support this because a temporary table
|
|
|
|
is always represented by only one TABLE object in THD, and
|
|
|
|
it can not be cloned. Emit an error for an unsupported behaviour.
|
|
|
|
*/
|
2007-11-01 21:52:56 +01:00
|
|
|
if (table->query_id)
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2007-02-23 12:13:55 +01:00
|
|
|
DBUG_PRINT("error",
|
|
|
|
("query_id: %lu server_id: %u pseudo_thread_id: %lu",
|
|
|
|
(ulong) table->query_id, (uint) thd->server_id,
|
|
|
|
(ulong) thd->variables.pseudo_thread_id));
|
2005-01-06 12:00:13 +01:00
|
|
|
my_error(ER_CANT_REOPEN_TABLE, MYF(0), table->alias);
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
|
|
|
table->query_id= thd->query_id;
|
2007-08-01 16:20:25 +02:00
|
|
|
thd->thread_specific_used= TRUE;
|
2004-10-29 18:26:52 +02:00
|
|
|
DBUG_PRINT("info",("Using temporary table"));
|
|
|
|
goto reset;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2007-05-11 19:51:03 +02:00
|
|
|
if (flags & MYSQL_OPEN_TEMPORARY_ONLY)
|
|
|
|
{
|
|
|
|
my_error(ER_NO_SUCH_TABLE, MYF(0), table_list->db, table_list->table_name);
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2007-05-11 19:51:03 +02:00
|
|
|
}
|
|
|
|
|
2007-02-28 14:03:38 +01:00
|
|
|
/*
|
|
|
|
The table is not temporary - if we're in pre-locked or LOCK TABLES
|
|
|
|
mode, let's try to find the requested table in the list of pre-opened
|
|
|
|
and locked tables. If the table is not there, return an error - we can't
|
|
|
|
open not pre-opened tables in pre-locked/LOCK TABLES mode.
|
|
|
|
TODO: move this block into a separate function.
|
|
|
|
*/
|
2009-12-02 16:22:15 +01:00
|
|
|
if (thd->locked_tables_mode &&
|
|
|
|
! (flags & MYSQL_OPEN_GET_NEW_TABLE))
|
2000-07-31 21:29:14 +02:00
|
|
|
{ // Using table locks
|
2005-07-15 00:22:14 +02:00
|
|
|
TABLE *best_table= 0;
|
2005-07-28 15:10:14 +02:00
|
|
|
int best_distance= INT_MIN;
|
2000-07-31 21:29:14 +02:00
|
|
|
for (table=thd->open_tables; table ; table=table->next)
|
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
if (table->s->table_cache_key.length == key_length &&
|
|
|
|
!memcmp(table->s->table_cache_key.str, key, key_length))
|
2004-06-25 15:52:01 +02:00
|
|
|
{
|
2005-08-18 17:07:23 +02:00
|
|
|
if (!my_strcasecmp(system_charset_info, table->alias, alias) &&
|
|
|
|
table->query_id != thd->query_id && /* skip tables already used */
|
2009-12-01 15:39:03 +01:00
|
|
|
(thd->locked_tables_mode == LTM_LOCK_TABLES ||
|
2009-12-03 00:09:22 +01:00
|
|
|
table->query_id == 0))
|
2005-08-18 17:07:23 +02:00
|
|
|
{
|
|
|
|
int distance= ((int) table->reginfo.lock_type -
|
|
|
|
(int) table_list->lock_type);
|
|
|
|
/*
|
|
|
|
Find a table that either has the exact lock type requested,
|
|
|
|
or has the best suitable lock. In case there is no locked
|
|
|
|
table that has an equal or higher lock than requested,
|
|
|
|
we us the closest matching lock to be able to produce an error
|
|
|
|
message about wrong lock mode on the table. The best_table
|
|
|
|
is changed if bd < 0 <= d or bd < d < 0 or 0 <= d < bd.
|
|
|
|
|
|
|
|
distance < 0 - No suitable lock found
|
|
|
|
distance > 0 - we have lock mode higher then we require
|
|
|
|
distance == 0 - we have lock mode exactly which we need
|
|
|
|
*/
|
2009-06-10 16:04:07 +02:00
|
|
|
if ((best_distance < 0 && distance > best_distance) ||
|
|
|
|
(distance >= 0 && distance < best_distance))
|
2005-08-18 17:07:23 +02:00
|
|
|
{
|
|
|
|
best_distance= distance;
|
|
|
|
best_table= table;
|
2009-03-27 12:09:15 +01:00
|
|
|
if (best_distance == 0)
|
2005-08-18 17:07:23 +02:00
|
|
|
{
|
|
|
|
/*
|
2009-03-27 12:09:15 +01:00
|
|
|
We have found a perfect match and can finish iterating
|
|
|
|
through open tables list. Check for table use conflict
|
|
|
|
between calling statement and SP/trigger is done in
|
|
|
|
lock_tables().
|
2005-08-18 17:07:23 +02:00
|
|
|
*/
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
2005-07-15 00:22:14 +02:00
|
|
|
}
|
2004-06-25 15:52:01 +02:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2005-07-15 00:22:14 +02:00
|
|
|
if (best_table)
|
|
|
|
{
|
|
|
|
table= best_table;
|
|
|
|
table->query_id= thd->query_id;
|
|
|
|
DBUG_PRINT("info",("Using locked table"));
|
|
|
|
goto reset;
|
|
|
|
}
|
2004-09-24 11:50:10 +02:00
|
|
|
/*
|
2007-02-28 14:03:38 +01:00
|
|
|
Is this table a view and not a base table?
|
2004-09-24 11:50:10 +02:00
|
|
|
(it is work around to allow to open view with locked tables,
|
|
|
|
real fix will be made after definition cache will be made)
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
Since opening of view which was not explicitly locked by LOCK
|
|
|
|
TABLES breaks metadata locking protocol (potentially can lead
|
|
|
|
to deadlocks) it should be disallowed.
|
2004-09-24 11:50:10 +02:00
|
|
|
*/
|
2009-12-10 09:21:38 +01:00
|
|
|
if (thd->mdl_context.is_lock_owner(MDL_key::TABLE, table_list->db,
|
2009-12-04 00:52:05 +01:00
|
|
|
table_list->table_name))
|
2004-09-24 11:50:10 +02:00
|
|
|
{
|
2009-06-19 10:24:43 +02:00
|
|
|
char path[FN_REFLEN + 1];
|
2005-12-21 19:18:40 +01:00
|
|
|
enum legacy_db_type not_used;
|
2005-12-31 06:01:26 +01:00
|
|
|
build_table_filename(path, sizeof(path) - 1,
|
2006-08-02 17:57:06 +02:00
|
|
|
table_list->db, table_list->table_name, reg_ext, 0);
|
2009-11-30 23:13:06 +01:00
|
|
|
/*
|
|
|
|
Note that we can't be 100% sure that it is a view since it's
|
|
|
|
possible that we either simply have not found unused TABLE
|
|
|
|
instance in THD::open_tables list or were unable to open table
|
|
|
|
during prelocking process (in this case in theory we still
|
|
|
|
should hold shared metadata lock on it).
|
|
|
|
*/
|
2005-11-03 15:10:11 +01:00
|
|
|
if (mysql_frm_type(thd, path, ¬_used) == FRMTYPE_VIEW)
|
2004-09-24 11:50:10 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
if (!tdc_open_view(thd, table_list, alias, key, key_length,
|
|
|
|
mem_root, 0))
|
2004-09-24 11:50:10 +02:00
|
|
|
{
|
2005-02-25 15:53:22 +01:00
|
|
|
DBUG_ASSERT(table_list->view != 0);
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(FALSE); // VIEW
|
2004-09-24 11:50:10 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2007-05-18 10:29:06 +02:00
|
|
|
/*
|
|
|
|
No table in the locked tables list. In case of explicit LOCK TABLES
|
|
|
|
this can happen if a user did not include the able into the list.
|
|
|
|
In case of pre-locked mode locked tables list is generated automatically,
|
|
|
|
so we may only end up here if the table did not exist when
|
|
|
|
locked tables list was created.
|
|
|
|
*/
|
2009-12-01 15:39:03 +01:00
|
|
|
if (thd->locked_tables_mode == LTM_PRELOCKED)
|
2009-12-02 16:22:15 +01:00
|
|
|
my_error(ER_NO_SUCH_TABLE, MYF(0), table_list->db, table_list->alias);
|
|
|
|
else
|
|
|
|
my_error(ER_TABLE_NOT_LOCKED, MYF(0), alias);
|
|
|
|
DBUG_RETURN(TRUE);
|
|
|
|
}
|
2009-12-01 14:59:11 +01:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
/*
|
|
|
|
Non pre-locked/LOCK TABLES mode, and the table is not temporary.
|
|
|
|
This is the normal use case.
|
|
|
|
*/
|
2009-12-01 14:59:11 +01:00
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
mdl_request= &table_list->mdl_request;
|
2009-12-02 16:22:15 +01:00
|
|
|
if (! (flags & MYSQL_OPEN_HAS_MDL_LOCK))
|
|
|
|
{
|
2009-12-08 10:57:07 +01:00
|
|
|
if (open_table_get_mdl_lock(thd, table_list, mdl_request, ot_ctx, flags))
|
|
|
|
{
|
|
|
|
DEBUG_SYNC(thd, "before_open_table_wait_refresh");
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2009-12-08 10:57:07 +01:00
|
|
|
}
|
|
|
|
DEBUG_SYNC(thd, "after_open_table_mdl_shared");
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
|
|
|
|
2009-12-04 00:29:40 +01:00
|
|
|
/*
|
|
|
|
Grab reference to the granted MDL lock ticket. Must be done after
|
|
|
|
open_table_get_mdl_lock as the lock on the table might have been
|
|
|
|
acquired previously (MYSQL_OPEN_HAS_MDL_LOCK).
|
|
|
|
*/
|
2009-12-04 00:52:05 +01:00
|
|
|
mdl_ticket= mdl_request->ticket;
|
2009-12-04 00:29:40 +01:00
|
|
|
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_lock(&LOCK_open);
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2007-02-28 14:03:38 +01:00
|
|
|
/*
|
|
|
|
If it's the first table from a list of tables used in a query,
|
|
|
|
remember refresh_version (the version of open_cache state).
|
|
|
|
If the version changes while we're opening the remaining tables,
|
|
|
|
we will have to back off, close all the tables opened-so-far,
|
|
|
|
and try to reopen them.
|
|
|
|
Note: refresh_version is currently changed only during FLUSH TABLES.
|
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
if (!thd->open_tables)
|
|
|
|
thd->version=refresh_version;
|
2005-09-13 14:36:43 +02:00
|
|
|
else if ((thd->version != refresh_version) &&
|
2006-03-29 13:27:36 +02:00
|
|
|
! (flags & MYSQL_LOCK_IGNORE_FLUSH))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
/* Someone did a refresh while thread was opening tables */
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2009-12-08 10:57:07 +01:00
|
|
|
(void) ot_ctx->request_backoff_action(Open_table_context::OT_WAIT);
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
Backport of revno ## 2617.31.1, 2617.31.3, 2617.31.4, 2617.31.5,
2617.31.12, 2617.31.15, 2617.31.15, 2617.31.16, 2617.43.1
- initial changeset that introduced the fix for
Bug#989 and follow up fixes for all test suite failures
introduced in the initial changeset.
------------------------------------------------------------
revno: 2617.31.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Fri 2009-03-06 19:17:00 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
Currently the MySQL server does not keep metadata locks on
schema objects for the duration of a transaction, thus failing
to guarantee the integrity of the schema objects being used
during the transaction and to protect then from concurrent
DDL operations. This also poses a problem for replication as
a DDL operation might be replicated even thought there are
active transactions using the object being modified.
The solution is to defer the release of metadata locks until
a active transaction is either committed or rolled back. This
prevents other statements from modifying the table for the
entire duration of the transaction. This provides commitment
ordering for guaranteeing serializability across multiple
transactions.
- Incompatible change:
If MySQL's metadata locking system encounters a lock conflict,
the usual schema is to use the try and back-off technique to
avoid deadlocks -- this schema consists in releasing all locks
and trying to acquire them all in one go.
But in a transactional context this algorithm can't be utilized
as its not possible to release locks acquired during the course
of the transaction without breaking the transaction commitments.
To avoid deadlocks in this case, the ER_LOCK_DEADLOCK will be
returned if a lock conflict is encountered during a transaction.
Let's consider an example:
A transaction has two statements that modify table t1, then table
t2, and then commits. The first statement of the transaction will
acquire a shared metadata lock on table t1, and it will be kept
utill COMMIT to ensure serializability.
At the moment when the second statement attempts to acquire a
shared metadata lock on t2, a concurrent ALTER or DROP statement
might have locked t2 exclusively. The prescription of the current
locking protocol is that the acquirer of the shared lock backs off
-- gives up all his current locks and retries. This implies that
the entire multi-statement transaction has to be rolled back.
- Incompatible change:
FLUSH commands such as FLUSH PRIVILEGES and FLUSH TABLES WITH READ
LOCK won't cause locked tables to be implicitly unlocked anymore.
2009-12-05 00:02:48 +01:00
|
|
|
if (table_list->open_strategy == TABLE_LIST::OPEN_IF_EXISTS)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
bool exists;
|
2006-05-30 15:07:49 +02:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
if (check_if_table_exists(thd, table_list, &exists))
|
|
|
|
goto err_unlock2;
|
2005-07-13 11:48:13 +02:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
if (!exists)
|
|
|
|
{
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(FALSE);
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
|
|
|
/* Table exists. Let us try to open it. */
|
|
|
|
}
|
Backport of revno ## 2617.31.1, 2617.31.3, 2617.31.4, 2617.31.5,
2617.31.12, 2617.31.15, 2617.31.15, 2617.31.16, 2617.43.1
- initial changeset that introduced the fix for
Bug#989 and follow up fixes for all test suite failures
introduced in the initial changeset.
------------------------------------------------------------
revno: 2617.31.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Fri 2009-03-06 19:17:00 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
Currently the MySQL server does not keep metadata locks on
schema objects for the duration of a transaction, thus failing
to guarantee the integrity of the schema objects being used
during the transaction and to protect then from concurrent
DDL operations. This also poses a problem for replication as
a DDL operation might be replicated even thought there are
active transactions using the object being modified.
The solution is to defer the release of metadata locks until
a active transaction is either committed or rolled back. This
prevents other statements from modifying the table for the
entire duration of the transaction. This provides commitment
ordering for guaranteeing serializability across multiple
transactions.
- Incompatible change:
If MySQL's metadata locking system encounters a lock conflict,
the usual schema is to use the try and back-off technique to
avoid deadlocks -- this schema consists in releasing all locks
and trying to acquire them all in one go.
But in a transactional context this algorithm can't be utilized
as its not possible to release locks acquired during the course
of the transaction without breaking the transaction commitments.
To avoid deadlocks in this case, the ER_LOCK_DEADLOCK will be
returned if a lock conflict is encountered during a transaction.
Let's consider an example:
A transaction has two statements that modify table t1, then table
t2, and then commits. The first statement of the transaction will
acquire a shared metadata lock on table t1, and it will be kept
utill COMMIT to ensure serializability.
At the moment when the second statement attempts to acquire a
shared metadata lock on t2, a concurrent ALTER or DROP statement
might have locked t2 exclusively. The prescription of the current
locking protocol is that the acquirer of the shared lock backs off
-- gives up all his current locks and retries. This implies that
the entire multi-statement transaction has to be rolled back.
- Incompatible change:
FLUSH commands such as FLUSH PRIVILEGES and FLUSH TABLES WITH READ
LOCK won't cause locked tables to be implicitly unlocked anymore.
2009-12-05 00:02:48 +01:00
|
|
|
else if (table_list->open_strategy == TABLE_LIST::OPEN_STUB)
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(FALSE);
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
2007-05-11 19:51:03 +02:00
|
|
|
|
2009-12-09 16:56:34 +01:00
|
|
|
#ifdef DISABLED_UNTIL_GRL_IS_MADE_PART_OF_MDL
|
2009-12-04 00:52:05 +01:00
|
|
|
if (!(share= (TABLE_SHARE *) mdl_ticket->get_cached_object()))
|
2009-12-09 16:56:34 +01:00
|
|
|
#endif
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
|
|
|
if (!(share= get_table_share_with_create(thd, table_list, key,
|
|
|
|
key_length, OPEN_VIEW,
|
|
|
|
&error)))
|
|
|
|
goto err_unlock2;
|
|
|
|
|
|
|
|
if (share->is_view)
|
|
|
|
{
|
2009-12-03 00:09:22 +01:00
|
|
|
/*
|
|
|
|
If parent_l of the table_list is non null then a merge table
|
|
|
|
has this view as child table, which is not supported.
|
|
|
|
*/
|
|
|
|
if (table_list->parent_l)
|
|
|
|
{
|
|
|
|
my_error(ER_WRONG_MRG_TABLE, MYF(0));
|
|
|
|
goto err_unlock;
|
|
|
|
}
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
/*
|
2009-11-30 16:55:03 +01:00
|
|
|
This table is a view. Validate its metadata version: in particular,
|
|
|
|
that it was a view when the statement was prepared.
|
2007-02-28 14:03:38 +01:00
|
|
|
*/
|
2009-11-30 16:55:03 +01:00
|
|
|
if (check_and_update_table_version(thd, table_list, share))
|
|
|
|
goto err_unlock;
|
|
|
|
if (table_list->i_s_requested_object & OPEN_TABLE_ONLY)
|
|
|
|
goto err_unlock;
|
|
|
|
|
|
|
|
/* Open view */
|
|
|
|
if (open_new_frm(thd, share, alias,
|
|
|
|
(uint) (HA_OPEN_KEYFILE | HA_OPEN_RNDFILE |
|
|
|
|
HA_GET_INDEX | HA_TRY_READ_ONLY),
|
|
|
|
READ_KEYINFO | COMPUTE_TYPES | EXTRA_RECORD |
|
|
|
|
(flags & OPEN_VIEW_NO_PARSE), thd->open_options,
|
|
|
|
0, table_list, mem_root))
|
|
|
|
goto err_unlock;
|
|
|
|
|
|
|
|
/* TODO: Don't free this */
|
2009-11-30 20:11:32 +01:00
|
|
|
release_table_share(share);
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
if (flags & OPEN_VIEW_NO_PARSE)
|
2007-02-28 14:03:38 +01:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
/*
|
|
|
|
VIEW not really opened, only frm were read.
|
|
|
|
Set 1 as a flag here
|
|
|
|
*/
|
|
|
|
table_list->view= (LEX*)1;
|
2007-02-28 14:03:38 +01:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
else
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
DBUG_ASSERT(table_list->view);
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(FALSE);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2009-12-01 15:58:31 +01:00
|
|
|
/*
|
|
|
|
Note that situation when we are trying to open a table for what
|
|
|
|
was a view during previous execution of PS will be handled in by
|
|
|
|
the caller. Here we should simply open our table even if
|
|
|
|
TABLE_LIST::view is true.
|
|
|
|
*/
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
if (table_list->i_s_requested_object & OPEN_VIEW_ONLY)
|
|
|
|
goto err_unlock;
|
|
|
|
|
2009-12-09 16:56:34 +01:00
|
|
|
#ifdef DISABLED_UNTIL_GRL_IS_MADE_PART_OF_MDL
|
2009-11-30 16:55:03 +01:00
|
|
|
/*
|
|
|
|
We are going to to store extra reference to the share in MDL-subsystem
|
|
|
|
so we need to increase reference counter;
|
|
|
|
*/
|
|
|
|
reference_table_share(share);
|
2009-12-04 00:52:05 +01:00
|
|
|
mdl_ticket->set_cached_object(share, table_share_release_hook);
|
2009-12-09 16:56:34 +01:00
|
|
|
#endif
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2009-12-09 16:56:34 +01:00
|
|
|
#ifdef DISABLED_UNTIL_GRL_IS_MADE_PART_OF_MDL
|
2009-11-30 16:55:03 +01:00
|
|
|
else
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
if (table_list->view)
|
|
|
|
{
|
|
|
|
DBUG_ASSERT(thd->m_reprepare_observer);
|
|
|
|
check_and_update_table_version(thd, table_list, share);
|
|
|
|
/* Always an error. */
|
|
|
|
DBUG_ASSERT(thd->is_error());
|
|
|
|
goto err_unlock;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
/* When we have cached TABLE_SHARE we know that is not a view. */
|
|
|
|
if (table_list->i_s_requested_object & OPEN_VIEW_ONLY)
|
|
|
|
goto err_unlock;
|
|
|
|
|
|
|
|
/*
|
|
|
|
We are going to use this share for construction of new TABLE object
|
|
|
|
so reference counter should be increased.
|
|
|
|
*/
|
|
|
|
reference_table_share(share);
|
|
|
|
}
|
2009-12-09 16:56:34 +01:00
|
|
|
#endif
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
if (share->version != refresh_version)
|
|
|
|
{
|
|
|
|
if (!(flags & MYSQL_LOCK_IGNORE_FLUSH))
|
|
|
|
{
|
2009-11-30 23:13:06 +01:00
|
|
|
/*
|
|
|
|
We already have an MDL lock. But we have encountered an old
|
|
|
|
version of table in the table definition cache which is possible
|
|
|
|
when someone changes the table version directly in the cache
|
|
|
|
without acquiring a metadata lock (e.g. this can happen during
|
|
|
|
"rolling" FLUSH TABLE(S)).
|
|
|
|
Note, that to avoid a "busywait" in this case, we have to wait
|
|
|
|
separately in the caller for old table versions to go away
|
|
|
|
(see tdc_wait_for_old_versions()).
|
|
|
|
*/
|
2009-11-30 20:11:32 +01:00
|
|
|
release_table_share(share);
|
2009-11-30 16:55:03 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2009-12-08 10:57:07 +01:00
|
|
|
(void) ot_ctx->request_backoff_action(Open_table_context::OT_WAIT);
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
|
|
|
/* Force close at once after usage */
|
|
|
|
thd->version= share->version;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!share->free_tables.is_empty())
|
|
|
|
{
|
|
|
|
table= share->free_tables.head();
|
|
|
|
table_def_use_table(thd, table);
|
|
|
|
/* We need to release share as we have EXTRA reference to it in our hands. */
|
2009-11-30 20:11:32 +01:00
|
|
|
release_table_share(share);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
/* We have too many TABLE instances around let us try to get rid of them. */
|
|
|
|
while (table_cache_count > table_cache_size && unused_tables)
|
|
|
|
free_cache_entry(unused_tables);
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/* make a new table */
|
|
|
|
if (!(table=(TABLE*) my_malloc(sizeof(*table),MYF(MY_WME))))
|
|
|
|
goto err_unlock;
|
2007-05-11 19:51:03 +02:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
error= open_table_from_share(thd, share, alias,
|
|
|
|
(uint) (HA_OPEN_KEYFILE |
|
|
|
|
HA_OPEN_RNDFILE |
|
|
|
|
HA_GET_INDEX |
|
|
|
|
HA_TRY_READ_ONLY),
|
|
|
|
(READ_KEYINFO | COMPUTE_TYPES |
|
|
|
|
EXTRA_RECORD),
|
|
|
|
thd->open_options, table, FALSE);
|
2007-05-11 19:51:03 +02:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
if (error)
|
|
|
|
{
|
|
|
|
my_free(table, MYF(0));
|
|
|
|
|
2009-11-30 23:13:06 +01:00
|
|
|
if (error == 7)
|
2007-05-11 19:51:03 +02:00
|
|
|
{
|
2009-11-30 23:13:06 +01:00
|
|
|
share->version= 0;
|
2009-12-08 10:57:07 +01:00
|
|
|
(void) ot_ctx->request_backoff_action(Open_table_context::OT_DISCOVER);
|
2009-11-30 23:13:06 +01:00
|
|
|
}
|
|
|
|
else if (share->crashed)
|
|
|
|
{
|
|
|
|
share->version= 0;
|
2009-12-08 10:57:07 +01:00
|
|
|
(void) ot_ctx->request_backoff_action(Open_table_context::OT_REPAIR);
|
2007-05-11 19:51:03 +02:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
goto err_unlock;
|
2002-08-05 17:50:38 +02:00
|
|
|
}
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
if (open_table_entry_fini(thd, share, table))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-11-30 16:55:03 +01:00
|
|
|
closefrm(table, 0);
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
my_free((uchar*)table, MYF(0));
|
2009-11-30 16:55:03 +01:00
|
|
|
goto err_unlock;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2006-08-08 22:05:42 +02:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/* Add table to the share's used tables list. */
|
|
|
|
table_def_add_used_table(thd, table);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2009-11-24 14:54:59 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2009-11-30 16:55:03 +01:00
|
|
|
|
2009-11-30 23:13:06 +01:00
|
|
|
/*
|
|
|
|
In CREATE TABLE .. If NOT EXISTS .. SELECT we have found that
|
|
|
|
table exists now we should downgrade our exclusive metadata
|
|
|
|
lock on this table to shared metadata lock.
|
|
|
|
*/
|
Backport of revno ## 2617.31.1, 2617.31.3, 2617.31.4, 2617.31.5,
2617.31.12, 2617.31.15, 2617.31.15, 2617.31.16, 2617.43.1
- initial changeset that introduced the fix for
Bug#989 and follow up fixes for all test suite failures
introduced in the initial changeset.
------------------------------------------------------------
revno: 2617.31.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Fri 2009-03-06 19:17:00 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
Currently the MySQL server does not keep metadata locks on
schema objects for the duration of a transaction, thus failing
to guarantee the integrity of the schema objects being used
during the transaction and to protect then from concurrent
DDL operations. This also poses a problem for replication as
a DDL operation might be replicated even thought there are
active transactions using the object being modified.
The solution is to defer the release of metadata locks until
a active transaction is either committed or rolled back. This
prevents other statements from modifying the table for the
entire duration of the transaction. This provides commitment
ordering for guaranteeing serializability across multiple
transactions.
- Incompatible change:
If MySQL's metadata locking system encounters a lock conflict,
the usual schema is to use the try and back-off technique to
avoid deadlocks -- this schema consists in releasing all locks
and trying to acquire them all in one go.
But in a transactional context this algorithm can't be utilized
as its not possible to release locks acquired during the course
of the transaction without breaking the transaction commitments.
To avoid deadlocks in this case, the ER_LOCK_DEADLOCK will be
returned if a lock conflict is encountered during a transaction.
Let's consider an example:
A transaction has two statements that modify table t1, then table
t2, and then commits. The first statement of the transaction will
acquire a shared metadata lock on table t1, and it will be kept
utill COMMIT to ensure serializability.
At the moment when the second statement attempts to acquire a
shared metadata lock on t2, a concurrent ALTER or DROP statement
might have locked t2 exclusively. The prescription of the current
locking protocol is that the acquirer of the shared lock backs off
-- gives up all his current locks and retries. This implies that
the entire multi-statement transaction has to be rolled back.
- Incompatible change:
FLUSH commands such as FLUSH PRIVILEGES and FLUSH TABLES WITH READ
LOCK won't cause locked tables to be implicitly unlocked anymore.
2009-12-05 00:02:48 +01:00
|
|
|
if (table_list->lock_strategy == TABLE_LIST::EXCLUSIVE_DOWNGRADABLE_MDL)
|
2009-12-04 00:52:05 +01:00
|
|
|
mdl_ticket->downgrade_exclusive_lock();
|
2009-11-30 16:55:03 +01:00
|
|
|
|
2009-12-04 00:52:05 +01:00
|
|
|
table->mdl_ticket= mdl_ticket;
|
2009-11-30 23:13:06 +01:00
|
|
|
|
|
|
|
table->next=thd->open_tables; /* Link into simple list */
|
|
|
|
thd->open_tables=table;
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
table->reginfo.lock_type=TL_READ; /* Assume read */
|
|
|
|
|
|
|
|
reset:
|
2005-11-23 21:45:02 +01:00
|
|
|
DBUG_ASSERT(table->s->ref_count > 0 || table->s->tmp_table != NO_TMP_TABLE);
|
|
|
|
|
2004-11-24 18:48:30 +01:00
|
|
|
if (thd->lex->need_correct_ident())
|
|
|
|
table->alias_name_used= my_strcasecmp(table_alias_charset,
|
2005-11-23 21:45:02 +01:00
|
|
|
table->s->table_name.str, alias);
|
2000-07-31 21:29:14 +02:00
|
|
|
/* Fix alias if table name changes */
|
2005-01-06 12:00:13 +01:00
|
|
|
if (strcmp(table->alias, alias))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2000-08-21 23:18:32 +02:00
|
|
|
uint length=(uint) strlen(alias)+1;
|
2005-01-06 12:00:13 +01:00
|
|
|
table->alias= (char*) my_realloc((char*) table->alias, length,
|
|
|
|
MYF(MY_WME));
|
|
|
|
memcpy((char*) table->alias, alias, length);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
table->tablenr=thd->current_tablenr++;
|
|
|
|
table->used_fields=0;
|
|
|
|
table->const_table=0;
|
2009-10-07 17:03:42 +02:00
|
|
|
table->null_row= table->maybe_null= 0;
|
|
|
|
table->force_index= table->force_index_order= table->force_index_group= 0;
|
2000-07-31 21:29:14 +02:00
|
|
|
table->status=STATUS_NO_RECORD;
|
2005-01-03 12:56:23 +01:00
|
|
|
table->insert_values= 0;
|
2005-09-22 02:23:07 +02:00
|
|
|
table->fulltext_searched= 0;
|
|
|
|
table->file->ft_handler= 0;
|
2009-06-26 21:57:42 +02:00
|
|
|
table->reginfo.impossible_range= 0;
|
2007-03-30 16:13:33 +02:00
|
|
|
/* Catch wrong handling of the auto_increment_field_not_null. */
|
|
|
|
DBUG_ASSERT(!table->auto_increment_field_not_null);
|
|
|
|
table->auto_increment_field_not_null= FALSE;
|
2004-06-25 15:52:01 +02:00
|
|
|
if (table->timestamp_field)
|
2004-10-01 16:54:06 +02:00
|
|
|
table->timestamp_field_type= table->timestamp_field->get_auto_set_type();
|
2006-11-01 02:31:56 +01:00
|
|
|
table->pos_in_table_list= table_list;
|
2004-07-16 00:15:55 +02:00
|
|
|
table_list->updatable= 1; // It is not derived table nor non-updatable VIEW
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
table->clear_column_bitmaps();
|
2009-12-01 15:58:31 +01:00
|
|
|
table_list->table= table;
|
2002-11-18 22:11:45 +01:00
|
|
|
DBUG_ASSERT(table->key_read == 0);
|
2009-12-03 00:09:22 +01:00
|
|
|
/* Tables may be reused in a sub statement. */
|
|
|
|
if (table->file->extra(HA_EXTRA_IS_ATTACHED_CHILDREN))
|
|
|
|
table->file->extra(HA_EXTRA_DETACH_CHILDREN);
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(FALSE);
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
err_unlock:
|
2009-11-30 20:11:32 +01:00
|
|
|
release_table_share(share);
|
2009-11-30 16:55:03 +01:00
|
|
|
err_unlock2:
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2009-12-02 17:15:40 +01:00
|
|
|
if (! (flags & MYSQL_OPEN_HAS_MDL_LOCK))
|
2009-12-04 00:52:05 +01:00
|
|
|
thd->mdl_context.release_lock(mdl_ticket);
|
2009-12-08 10:57:07 +01:00
|
|
|
|
2009-12-01 15:58:31 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/**
|
|
|
|
Find table in the list of open tables.
|
|
|
|
|
|
|
|
@param list List of TABLE objects to be inspected.
|
|
|
|
@param db Database name
|
|
|
|
@param table_name Table name
|
|
|
|
|
|
|
|
@return Pointer to the TABLE object found, 0 if no table found.
|
|
|
|
*/
|
|
|
|
|
|
|
|
TABLE *find_locked_table(TABLE *list, const char *db, const char *table_name)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
char key[MAX_DBKEY_LENGTH];
|
|
|
|
uint key_length=(uint) (strmov(strmov(key,db)+1,table_name)-key)+1;
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
for (TABLE *table= list; table ; table=table->next)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
if (table->s->table_cache_key.length == key_length &&
|
|
|
|
!memcmp(table->s->table_cache_key.str, key, key_length))
|
2000-07-31 21:29:14 +02:00
|
|
|
return table;
|
|
|
|
}
|
|
|
|
return(0);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/**
|
|
|
|
Find write locked instance of table in the list of open tables,
|
|
|
|
emit error if no such instance found.
|
|
|
|
|
|
|
|
@param thd List of TABLE objects to be searched
|
|
|
|
@param db Database name.
|
|
|
|
@param table_name Name of table.
|
|
|
|
|
|
|
|
@return Pointer to write-locked TABLE instance, 0 - otherwise.
|
|
|
|
*/
|
|
|
|
|
|
|
|
TABLE *find_write_locked_table(TABLE *list, const char *db, const char *table_name)
|
|
|
|
{
|
|
|
|
TABLE *tab= find_locked_table(list, db, table_name);
|
|
|
|
|
|
|
|
if (!tab)
|
|
|
|
{
|
|
|
|
my_error(ER_TABLE_NOT_LOCKED, MYF(0), table_name);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
while (tab->reginfo.lock_type < TL_WRITE_LOW_PRIORITY &&
|
|
|
|
(tab= find_locked_table(tab->next, db, table_name)))
|
|
|
|
continue;
|
|
|
|
if (!tab)
|
|
|
|
{
|
|
|
|
my_error(ER_TABLE_NOT_LOCKED_FOR_WRITE, MYF(0), table_name);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return tab;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-09 10:44:01 +01:00
|
|
|
/**
|
|
|
|
Find instance of TABLE with MDL_SHARED_UPGRADABLE or
|
|
|
|
MDL_EXCLUSIVE lock from the list of open tables.
|
|
|
|
|
|
|
|
@param list List of TABLE objects to be searched
|
|
|
|
@param db Database name.
|
|
|
|
@param table_name Name of table.
|
|
|
|
|
|
|
|
@return Pointer to MDL_SHARED_UPGRADABLE or MDL_EXCLUSIVE
|
|
|
|
TABLE instance, NULL otherwise.
|
|
|
|
*/
|
|
|
|
|
|
|
|
TABLE *find_table_for_mdl_upgrade(TABLE *list, const char *db,
|
|
|
|
const char *table_name)
|
|
|
|
{
|
|
|
|
TABLE *tab= find_locked_table(list, db, table_name);
|
|
|
|
|
|
|
|
while (tab != NULL)
|
|
|
|
{
|
|
|
|
if (tab->mdl_ticket != NULL &&
|
|
|
|
tab->mdl_ticket->is_upgradable_or_exclusive())
|
|
|
|
return tab;
|
|
|
|
tab= find_locked_table(tab->next, db, table_name);
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
/***********************************************************************
|
|
|
|
class Locked_tables_list implementation. Declared in sql_class.h
|
|
|
|
************************************************************************/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
/**
|
|
|
|
Enter LTM_LOCK_TABLES mode.
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
Enter the LOCK TABLES mode using all the tables that are
|
|
|
|
currently open and locked in this connection.
|
|
|
|
Initializes a TABLE_LIST instance for every locked table.
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
@param thd thread handle
|
|
|
|
|
|
|
|
@return TRUE if out of memory.
|
2005-11-23 21:45:02 +01:00
|
|
|
*/
|
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
bool
|
|
|
|
Locked_tables_list::init_locked_tables(THD *thd)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
DBUG_ASSERT(thd->locked_tables_mode == LTM_NONE);
|
|
|
|
DBUG_ASSERT(m_locked_tables == NULL);
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
DBUG_ASSERT(m_reopen_array == NULL);
|
|
|
|
DBUG_ASSERT(m_locked_tables_count == 0);
|
2009-12-02 16:22:15 +01:00
|
|
|
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
for (TABLE *table= thd->open_tables; table;
|
|
|
|
table= table->next, m_locked_tables_count++)
|
2009-12-02 16:22:15 +01:00
|
|
|
{
|
|
|
|
TABLE_LIST *src_table_list= table->pos_in_table_list;
|
|
|
|
char *db, *table_name, *alias;
|
2009-12-02 16:37:10 +01:00
|
|
|
size_t db_len= src_table_list->db_length;
|
|
|
|
size_t table_name_len= src_table_list->table_name_length;
|
|
|
|
size_t alias_len= strlen(src_table_list->alias);
|
2009-12-02 16:22:15 +01:00
|
|
|
TABLE_LIST *dst_table_list;
|
|
|
|
|
|
|
|
if (! multi_alloc_root(&m_locked_tables_root,
|
|
|
|
&dst_table_list, sizeof(*dst_table_list),
|
2009-12-02 16:37:10 +01:00
|
|
|
&db, db_len + 1,
|
|
|
|
&table_name, table_name_len + 1,
|
|
|
|
&alias, alias_len + 1,
|
2009-12-02 16:33:51 +01:00
|
|
|
NullS))
|
2009-12-02 16:22:15 +01:00
|
|
|
{
|
|
|
|
unlock_locked_tables(0);
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
memcpy(db, src_table_list->db, db_len + 1);
|
|
|
|
memcpy(table_name, src_table_list->table_name, table_name_len + 1);
|
|
|
|
memcpy(alias, src_table_list->alias, alias_len + 1);
|
2009-12-02 16:22:15 +01:00
|
|
|
/**
|
|
|
|
Sic: remember the *actual* table level lock type taken, to
|
|
|
|
acquire the exact same type in reopen_tables().
|
|
|
|
E.g. if the table was locked for write, src_table_list->lock_type is
|
|
|
|
TL_WRITE_DEFAULT, whereas reginfo.lock_type has been updated from
|
|
|
|
thd->update_lock_default.
|
|
|
|
*/
|
2009-12-02 16:37:10 +01:00
|
|
|
dst_table_list->init_one_table(db, db_len, table_name, table_name_len,
|
|
|
|
alias,
|
2009-12-02 16:22:15 +01:00
|
|
|
src_table_list->table->reginfo.lock_type);
|
|
|
|
dst_table_list->table= table;
|
2009-12-08 10:57:07 +01:00
|
|
|
dst_table_list->mdl_request.ticket= src_table_list->mdl_request.ticket;
|
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
/* Link last into the list of tables */
|
|
|
|
*(dst_table_list->prev_global= m_locked_tables_last)= dst_table_list;
|
|
|
|
m_locked_tables_last= &dst_table_list->next_global;
|
|
|
|
table->pos_in_locked_tables= dst_table_list;
|
|
|
|
}
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
if (m_locked_tables_count)
|
|
|
|
{
|
|
|
|
/**
|
|
|
|
Allocate an auxiliary array to pass to mysql_lock_tables()
|
|
|
|
in reopen_tables(). reopen_tables() is a critical
|
|
|
|
path and we don't want to complicate it with extra allocations.
|
|
|
|
*/
|
|
|
|
m_reopen_array= (TABLE**)alloc_root(&m_locked_tables_root,
|
|
|
|
sizeof(TABLE*) *
|
|
|
|
(m_locked_tables_count+1));
|
|
|
|
if (m_reopen_array == NULL)
|
|
|
|
{
|
|
|
|
unlock_locked_tables(0);
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
}
|
2009-12-02 16:22:15 +01:00
|
|
|
thd->locked_tables_mode= LTM_LOCK_TABLES;
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
return FALSE;
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
/**
|
|
|
|
Leave LTM_LOCK_TABLES mode if it's been entered.
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
Close all locked tables, free memory, and leave the mode.
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
@note This function is a no-op if we're not in LOCK TABLES.
|
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
void
|
|
|
|
Locked_tables_list::unlock_locked_tables(THD *thd)
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
{
|
|
|
|
if (thd)
|
2006-01-10 18:13:12 +01:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
DBUG_ASSERT(!thd->in_sub_stmt &&
|
|
|
|
!(thd->state_flags & Open_tables_state::BACKUPS_AVAIL));
|
|
|
|
/*
|
|
|
|
Sic: we must be careful to not close open tables if
|
|
|
|
we're not in LOCK TABLES mode: unlock_locked_tables() is
|
|
|
|
sometimes called implicitly, expecting no effect on
|
|
|
|
open tables, e.g. from begin_trans().
|
|
|
|
*/
|
|
|
|
if (thd->locked_tables_mode != LTM_LOCK_TABLES)
|
|
|
|
return;
|
|
|
|
|
|
|
|
for (TABLE_LIST *table_list= m_locked_tables;
|
|
|
|
table_list; table_list= table_list->next_global)
|
2008-11-27 15:54:23 +01:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
/*
|
|
|
|
Clear the position in the list, the TABLE object will be
|
|
|
|
returned to the table cache.
|
|
|
|
*/
|
|
|
|
table_list->table->pos_in_locked_tables= NULL;
|
2008-11-27 15:54:23 +01:00
|
|
|
}
|
2009-12-02 16:22:15 +01:00
|
|
|
thd->locked_tables_mode= LTM_NONE;
|
|
|
|
|
|
|
|
close_thread_tables(thd);
|
2006-01-10 18:13:12 +01:00
|
|
|
}
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
/*
|
2009-12-02 16:22:15 +01:00
|
|
|
After closing tables we can free memory used for storing lock
|
|
|
|
request for metadata locks and TABLE_LIST elements.
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
*/
|
2009-12-02 16:22:15 +01:00
|
|
|
free_root(&m_locked_tables_root, MYF(0));
|
|
|
|
m_locked_tables= NULL;
|
|
|
|
m_locked_tables_last= &m_locked_tables;
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
m_reopen_array= NULL;
|
|
|
|
m_locked_tables_count= 0;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-05-19 08:49:56 +02:00
|
|
|
/**
|
2009-12-02 16:22:15 +01:00
|
|
|
Unlink a locked table from the locked tables list, either
|
|
|
|
temporarily or permanently.
|
|
|
|
|
|
|
|
@param thd thread handle
|
|
|
|
@param table_list the element of locked tables list.
|
|
|
|
The implementation assumes that this argument
|
|
|
|
points to a TABLE_LIST element linked into
|
|
|
|
the locked tables list. Passing a TABLE_LIST
|
|
|
|
instance that is not part of locked tables
|
|
|
|
list will lead to a crash.
|
|
|
|
@parma remove_from_locked_tables
|
|
|
|
TRUE if the table is removed from the list
|
|
|
|
permanently.
|
|
|
|
|
|
|
|
This function is a no-op if we're not under LOCK TABLES.
|
|
|
|
|
|
|
|
@sa Locked_tables_list::reopen_tables()
|
2005-11-23 21:45:02 +01:00
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
|
|
|
|
void Locked_tables_list::unlink_from_list(THD *thd,
|
|
|
|
TABLE_LIST *table_list,
|
|
|
|
bool remove_from_locked_tables)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
/*
|
|
|
|
If mode is not LTM_LOCK_TABLES, we needn't do anything. Moreover,
|
|
|
|
outside this mode pos_in_locked_tables value is not trustworthy.
|
|
|
|
*/
|
|
|
|
if (thd->locked_tables_mode != LTM_LOCK_TABLES)
|
|
|
|
return;
|
2007-05-19 08:49:56 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
/*
|
|
|
|
table_list must be set and point to pos_in_locked_tables of some
|
|
|
|
table.
|
|
|
|
*/
|
|
|
|
DBUG_ASSERT(table_list->table->pos_in_locked_tables == table_list);
|
2007-05-19 08:49:56 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
/* Clear the pointer, the table will be returned to the table cache. */
|
|
|
|
table_list->table->pos_in_locked_tables= NULL;
|
|
|
|
|
|
|
|
/* Mark the table as closed in the locked tables list. */
|
|
|
|
table_list->table= NULL;
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
/*
|
|
|
|
If the table is being dropped or renamed, remove it from
|
|
|
|
the locked tables list (implicitly drop the LOCK TABLES lock
|
|
|
|
on it).
|
|
|
|
*/
|
|
|
|
if (remove_from_locked_tables)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
*table_list->prev_global= table_list->next_global;
|
|
|
|
if (table_list->next_global == NULL)
|
|
|
|
m_locked_tables_last= table_list->prev_global;
|
|
|
|
else
|
|
|
|
table_list->next_global->prev_global= table_list->prev_global;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2007-05-11 19:51:03 +02:00
|
|
|
/**
|
2009-12-02 16:22:15 +01:00
|
|
|
This is an attempt to recover (somewhat) in case of an error.
|
|
|
|
If we failed to reopen a closed table, let's unlink it from the
|
|
|
|
list and forget about it. From a user perspective that would look
|
|
|
|
as if the server "lost" the lock on one of the locked tables.
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
@note This function is a no-op if we're not under LOCK TABLES.
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
*/
|
|
|
|
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
void Locked_tables_list::
|
|
|
|
unlink_all_closed_tables(THD *thd, MYSQL_LOCK *lock, size_t reopen_count)
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
{
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
/* If we managed to take a lock, unlock tables and free the lock. */
|
|
|
|
if (lock)
|
|
|
|
mysql_unlock_tables(thd, lock);
|
|
|
|
/*
|
|
|
|
If a failure happened in reopen_tables(), we may have succeeded
|
|
|
|
reopening some tables, but not all.
|
|
|
|
This works when the connection was killed in mysql_lock_tables().
|
|
|
|
*/
|
|
|
|
if (reopen_count)
|
|
|
|
{
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
while (reopen_count--)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
When closing the table, we must remove it
|
|
|
|
from thd->open_tables list.
|
|
|
|
We rely on the fact that open_table() that was used
|
|
|
|
in reopen_tables() always links the opened table
|
|
|
|
to the beginning of the open_tables list.
|
|
|
|
*/
|
|
|
|
DBUG_ASSERT(thd->open_tables == m_reopen_array[reopen_count]);
|
|
|
|
|
|
|
|
thd->open_tables->pos_in_locked_tables->table= NULL;
|
|
|
|
|
|
|
|
close_thread_table(thd, &thd->open_tables);
|
|
|
|
}
|
|
|
|
broadcast_refresh();
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
}
|
|
|
|
/* Exclude all closed tables from the LOCK TABLES list. */
|
2009-12-02 16:22:15 +01:00
|
|
|
for (TABLE_LIST *table_list= m_locked_tables; table_list; table_list=
|
|
|
|
table_list->next_global)
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
if (table_list->table == NULL)
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
/* Unlink from list. */
|
|
|
|
*table_list->prev_global= table_list->next_global;
|
|
|
|
if (table_list->next_global == NULL)
|
|
|
|
m_locked_tables_last= table_list->prev_global;
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
else
|
2009-12-02 16:22:15 +01:00
|
|
|
table_list->next_global->prev_global= table_list->prev_global;
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-05-11 19:51:03 +02:00
|
|
|
/**
|
2009-12-02 16:22:15 +01:00
|
|
|
Reopen the tables locked with LOCK TABLES and temporarily closed
|
|
|
|
by a DDL statement or FLUSH TABLES.
|
2007-05-11 19:51:03 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
@note This function is a no-op if we're not under LOCK TABLES.
|
2007-05-11 19:51:03 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
@return TRUE if an error reopening the tables. May happen in
|
|
|
|
case of some fatal system error only, e.g. a disk
|
|
|
|
corruption, out of memory or a serious bug in the
|
|
|
|
locking.
|
2000-07-31 21:29:14 +02:00
|
|
|
*/
|
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
bool
|
|
|
|
Locked_tables_list::reopen_tables(THD *thd)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-12-08 10:57:07 +01:00
|
|
|
Open_table_context ot_ctx_unused(thd);
|
2009-12-02 16:22:15 +01:00
|
|
|
bool lt_refresh_unused;
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
size_t reopen_count= 0;
|
|
|
|
MYSQL_LOCK *lock;
|
|
|
|
MYSQL_LOCK *merged_lock;
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-12-02 16:22:15 +01:00
|
|
|
for (TABLE_LIST *table_list= m_locked_tables;
|
|
|
|
table_list; table_list= table_list->next_global)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
if (table_list->table) /* The table was not closed */
|
|
|
|
continue;
|
|
|
|
|
|
|
|
/* Links into thd->open_tables upon success */
|
2009-12-08 10:57:07 +01:00
|
|
|
if (open_table(thd, table_list, thd->mem_root, &ot_ctx_unused,
|
2009-12-02 16:22:15 +01:00
|
|
|
MYSQL_OPEN_REOPEN))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
unlink_all_closed_tables(thd, 0, reopen_count);
|
2009-12-02 16:22:15 +01:00
|
|
|
return TRUE;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2009-12-02 16:22:15 +01:00
|
|
|
table_list->table->pos_in_locked_tables= table_list;
|
|
|
|
/* See also the comment on lock type in init_locked_tables(). */
|
|
|
|
table_list->table->reginfo.lock_type= table_list->lock_type;
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
|
|
|
|
DBUG_ASSERT(reopen_count < m_locked_tables_count);
|
|
|
|
m_reopen_array[reopen_count++]= table_list->table;
|
|
|
|
}
|
|
|
|
if (reopen_count)
|
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
thd->in_lock_tables= 1;
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
/*
|
|
|
|
We re-lock all tables with mysql_lock_tables() at once rather
|
|
|
|
than locking one table at a time because of the case
|
|
|
|
reported in Bug#45035: when the same table is present
|
|
|
|
in the list many times, thr_lock.c fails to grant READ lock
|
|
|
|
on a table that is already locked by WRITE lock, even if
|
|
|
|
WRITE lock is taken by the same thread. If READ and WRITE
|
|
|
|
lock are passed to thr_lock.c in the same list, everything
|
|
|
|
works fine. Patching legacy code of thr_lock.c is risking to
|
|
|
|
break something else.
|
|
|
|
*/
|
|
|
|
lock= mysql_lock_tables(thd, m_reopen_array, reopen_count,
|
2009-12-02 16:22:15 +01:00
|
|
|
MYSQL_OPEN_REOPEN, <_refresh_unused);
|
|
|
|
thd->in_lock_tables= 0;
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
if (lock == NULL || (merged_lock=
|
|
|
|
mysql_lock_merge(thd->lock, lock)) == NULL)
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
{
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
unlink_all_closed_tables(thd, lock, reopen_count);
|
|
|
|
if (! thd->killed)
|
|
|
|
my_error(ER_LOCK_DEADLOCK, MYF(0));
|
2009-12-02 16:22:15 +01:00
|
|
|
return TRUE;
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
}
|
Backport of:
----------------------------------------------------------
revno: 2617.69.2
committer: Konstantin Osipov <kostja@sun.com>
branch nick: 5.4-azalea-bugfixing
timestamp: Mon 2009-08-03 19:26:04 +0400
message:
A fix and a test case for Bug#45035 "Altering table under LOCK TABLES
results in "Error 1213 Deadlock found...".
If a user had a table locked with LOCK TABLES
for READ and for WRITE in the same connection, ALTER TABLE
could fail.
Root cause analysis:
If a connection issues
LOCK TABLE t1 write, t1 a read, t1 b read;
the new LOCK TABLES code in 6.0 (part of WL 3726) will create
the following list of TABLE_LIST objects
(thd->locked_tables_list->m_locked_tables):
{"t1" "b" tl_read_no_insert}, {"t1" "a" tl_read_no_insert},
{"t1" "t1" tl_write }
Later on, when we try to ALTER table t1, mysql_alter_table()
closes all TABLE instances and releases its thr_lock locks,
keeping only an exclusive metadata lock on t1.
But when ALTER is finished, Locked_table_list::reopen_tables()
tries to restore the original list of open and locked tables.
Before this patch, it used to do so one by one:
Open t1 b, get TL_READ_NO_INSERT lock,
Open t1 a, get TL_READ_NO_INSERT lock
Open t1, try to get TL_WRITE lock, deadlock.
The cause of the deadlock is that thr_lock.c doesn't
resolve the situation when the read list only consists
of locks taken by the same thread, followed by this very
thread trying to take a WRITE lock. Indeed, since
thr_lock_multi always gets a sorted list of locks,
WRITE locks always precede READ locks in the list
to lock.
Don't try to fix thr_lock.c deficiency, keep this
code simple.
Instead, try to take all thr_lock locks at once
in ::reopen_tables().
2009-12-08 09:38:45 +01:00
|
|
|
thd->lock= merged_lock;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2009-12-02 16:22:15 +01:00
|
|
|
return FALSE;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2003-03-04 11:22:35 +01:00
|
|
|
|
2005-12-22 06:39:02 +01:00
|
|
|
/*
|
2006-02-23 13:34:03 +01:00
|
|
|
Function to assign a new table map id to a table share.
|
2005-12-22 06:39:02 +01:00
|
|
|
|
|
|
|
PARAMETERS
|
|
|
|
|
2006-02-23 13:34:03 +01:00
|
|
|
share - Pointer to table share structure
|
2005-12-22 06:39:02 +01:00
|
|
|
|
2006-02-27 17:23:20 +01:00
|
|
|
DESCRIPTION
|
|
|
|
|
|
|
|
We are intentionally not checking that share->mutex is locked
|
|
|
|
since this function should only be called when opening a table
|
|
|
|
share and before it is entered into the table_def_cache (meaning
|
|
|
|
that it cannot be fetched by another thread, even accidentally).
|
|
|
|
|
2005-12-22 06:39:02 +01:00
|
|
|
PRE-CONDITION(S)
|
|
|
|
|
2006-02-23 13:34:03 +01:00
|
|
|
share is non-NULL
|
2005-12-22 06:39:02 +01:00
|
|
|
The LOCK_open mutex is locked
|
|
|
|
|
|
|
|
POST-CONDITION(S)
|
|
|
|
|
2006-02-23 13:34:03 +01:00
|
|
|
share->table_map_id is given a value that with a high certainty is
|
|
|
|
not used by any other table (the only case where a table id can be
|
|
|
|
reused is on wrap-around, which means more than 4 billion table
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
share opens have been executed while one table was open all the
|
|
|
|
time).
|
2005-12-22 06:39:02 +01:00
|
|
|
|
2006-03-09 03:56:14 +01:00
|
|
|
share->table_map_id is not ~0UL.
|
2005-12-22 06:39:02 +01:00
|
|
|
*/
|
2008-08-20 16:06:31 +02:00
|
|
|
static ulong last_table_id= ~0UL;
|
|
|
|
|
2006-02-23 13:34:03 +01:00
|
|
|
void assign_new_table_id(TABLE_SHARE *share)
|
2005-12-22 06:39:02 +01:00
|
|
|
{
|
|
|
|
|
2006-02-23 13:34:03 +01:00
|
|
|
DBUG_ENTER("assign_new_table_id");
|
2005-12-22 06:39:02 +01:00
|
|
|
|
|
|
|
/* Preconditions */
|
2006-02-23 13:34:03 +01:00
|
|
|
DBUG_ASSERT(share != NULL);
|
2005-12-22 06:39:02 +01:00
|
|
|
safe_mutex_assert_owner(&LOCK_open);
|
|
|
|
|
|
|
|
ulong tid= ++last_table_id; /* get next id */
|
2006-02-23 13:34:03 +01:00
|
|
|
/*
|
|
|
|
There is one reserved number that cannot be used. Remember to
|
|
|
|
change this when 6-byte global table id's are introduced.
|
|
|
|
*/
|
2006-03-09 03:56:14 +01:00
|
|
|
if (unlikely(tid == ~0UL))
|
2005-12-22 06:39:02 +01:00
|
|
|
tid= ++last_table_id;
|
2006-02-23 13:34:03 +01:00
|
|
|
share->table_map_id= tid;
|
2005-12-22 06:39:02 +01:00
|
|
|
DBUG_PRINT("info", ("table_id=%lu", tid));
|
|
|
|
|
|
|
|
/* Post conditions */
|
2006-03-09 03:56:14 +01:00
|
|
|
DBUG_ASSERT(share->table_map_id != ~0UL);
|
2005-12-22 06:39:02 +01:00
|
|
|
|
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
}
|
|
|
|
|
2008-10-03 13:16:58 +02:00
|
|
|
#ifndef DBUG_OFF
|
2008-10-02 21:13:42 +02:00
|
|
|
/* Cause a spurious statement reprepare for debug purposes. */
|
2008-10-03 13:16:58 +02:00
|
|
|
static bool inject_reprepare(THD *thd)
|
2008-10-02 21:13:42 +02:00
|
|
|
{
|
|
|
|
if (thd->m_reprepare_observer && thd->stmt_arena->is_reprepared == FALSE)
|
|
|
|
{
|
|
|
|
thd->m_reprepare_observer->report_error(thd);
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
return FALSE;
|
|
|
|
}
|
2008-10-03 13:16:58 +02:00
|
|
|
#endif
|
2008-10-02 21:13:42 +02:00
|
|
|
|
2008-04-08 18:01:20 +02:00
|
|
|
/**
|
|
|
|
Compare metadata versions of an element obtained from the table
|
|
|
|
definition cache and its corresponding node in the parse tree.
|
|
|
|
|
2008-05-17 23:51:18 +02:00
|
|
|
@details If the new and the old values mismatch, invoke
|
2008-04-08 18:01:20 +02:00
|
|
|
Metadata_version_observer.
|
|
|
|
At prepared statement prepare, all TABLE_LIST version values are
|
|
|
|
NULL and we always have a mismatch. But there is no observer set
|
|
|
|
in THD, and therefore no error is reported. Instead, we update
|
|
|
|
the value in the parse tree, effectively recording the original
|
|
|
|
version.
|
|
|
|
At prepared statement execute, an observer may be installed. If
|
|
|
|
there is a version mismatch, we push an error and return TRUE.
|
|
|
|
|
|
|
|
For conventional execution (no prepared statements), the
|
|
|
|
observer is never installed.
|
|
|
|
|
|
|
|
@sa Execute_observer
|
|
|
|
@sa check_prepared_statement() to see cases when an observer is installed
|
2008-05-20 09:29:16 +02:00
|
|
|
@sa TABLE_LIST::is_table_ref_id_equal()
|
|
|
|
@sa TABLE_SHARE::get_table_ref_id()
|
2008-04-08 18:01:20 +02:00
|
|
|
|
|
|
|
@param[in] thd used to report errors
|
|
|
|
@param[in,out] tables TABLE_LIST instance created by the parser
|
|
|
|
Metadata version information in this object
|
|
|
|
is updated upon success.
|
|
|
|
@param[in] table_share an element from the table definition cache
|
|
|
|
|
|
|
|
@retval TRUE an error, which has been reported
|
|
|
|
@retval FALSE success, version in TABLE_LIST has been updated
|
|
|
|
*/
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
static bool
|
2008-04-08 18:01:20 +02:00
|
|
|
check_and_update_table_version(THD *thd,
|
|
|
|
TABLE_LIST *tables, TABLE_SHARE *table_share)
|
|
|
|
{
|
2008-05-20 09:29:16 +02:00
|
|
|
if (! tables->is_table_ref_id_equal(table_share))
|
2008-04-08 18:01:20 +02:00
|
|
|
{
|
2008-05-20 09:29:16 +02:00
|
|
|
if (thd->m_reprepare_observer &&
|
|
|
|
thd->m_reprepare_observer->report_error(thd))
|
2008-04-08 18:01:20 +02:00
|
|
|
{
|
|
|
|
/*
|
|
|
|
Version of the table share is different from the
|
|
|
|
previous execution of the prepared statement, and it is
|
|
|
|
unacceptable for this SQLCOM. Error has been reported.
|
|
|
|
*/
|
2008-05-17 23:51:18 +02:00
|
|
|
DBUG_ASSERT(thd->is_error());
|
2008-04-08 18:01:20 +02:00
|
|
|
return TRUE;
|
|
|
|
}
|
2008-05-17 23:51:18 +02:00
|
|
|
/* Always maintain the latest version and type */
|
2008-05-20 09:29:16 +02:00
|
|
|
tables->set_table_ref_id(table_share);
|
2008-04-08 18:01:20 +02:00
|
|
|
}
|
2008-05-19 18:39:31 +02:00
|
|
|
|
2008-10-02 21:13:42 +02:00
|
|
|
DBUG_EXECUTE_IF("reprepare_each_statement", return inject_reprepare(thd););
|
2008-05-19 18:39:31 +02:00
|
|
|
|
2008-04-08 18:01:20 +02:00
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
2003-03-04 11:22:35 +01:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/**
|
|
|
|
Open view by getting its definition from disk (and table cache in future).
|
2003-03-04 11:22:35 +01:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
@param thd Thread handle
|
|
|
|
@param table_list TABLE_LIST with db, table_name & belong_to_view
|
|
|
|
@param alias Alias name
|
|
|
|
@param cache_key Key for table definition cache
|
|
|
|
@param cache_key_length Length of cache_key
|
|
|
|
@param mem_root Memory to be used for .frm parsing.
|
|
|
|
@param flags Flags which modify how we open the view
|
2003-03-04 11:22:35 +01:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
@todo This function is needed for special handling of views under
|
|
|
|
LOCK TABLES. We probably should get rid of it in long term.
|
|
|
|
|
|
|
|
@return FALSE if success, TRUE - otherwise.
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool tdc_open_view(THD *thd, TABLE_LIST *table_list, const char *alias,
|
|
|
|
char *cache_key, uint cache_key_length,
|
|
|
|
MEM_ROOT *mem_root, uint flags)
|
|
|
|
{
|
|
|
|
TABLE not_used;
|
|
|
|
int error;
|
|
|
|
TABLE_SHARE *share;
|
|
|
|
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
|
|
|
|
if (!(share= get_table_share_with_create(thd, table_list, cache_key,
|
|
|
|
cache_key_length,
|
|
|
|
OPEN_VIEW, &error)))
|
|
|
|
goto err;
|
|
|
|
|
|
|
|
if (share->is_view &&
|
|
|
|
!open_new_frm(thd, share, alias,
|
|
|
|
(uint) (HA_OPEN_KEYFILE | HA_OPEN_RNDFILE |
|
|
|
|
HA_GET_INDEX | HA_TRY_READ_ONLY),
|
|
|
|
READ_KEYINFO | COMPUTE_TYPES | EXTRA_RECORD |
|
|
|
|
flags, thd->open_options, ¬_used, table_list,
|
|
|
|
mem_root))
|
|
|
|
{
|
2009-11-30 20:11:32 +01:00
|
|
|
release_table_share(share);
|
2009-11-30 16:55:03 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
my_error(ER_WRONG_OBJECT, MYF(0), share->db.str, share->table_name.str, "VIEW");
|
2009-11-30 20:11:32 +01:00
|
|
|
release_table_share(share);
|
2009-11-30 16:55:03 +01:00
|
|
|
err:
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
2009-11-30 23:13:06 +01:00
|
|
|
Finalize the process of TABLE creation by loading table triggers
|
|
|
|
and taking action if a HEAP table content was emptied implicitly.
|
2009-11-30 16:55:03 +01:00
|
|
|
*/
|
|
|
|
|
|
|
|
static bool open_table_entry_fini(THD *thd, TABLE_SHARE *share, TABLE *entry)
|
|
|
|
{
|
|
|
|
|
|
|
|
if (Table_triggers_list::check_n_load(thd, share->db.str,
|
|
|
|
share->table_name.str, entry, 0))
|
|
|
|
return TRUE;
|
|
|
|
|
2004-06-25 15:52:01 +02:00
|
|
|
/*
|
|
|
|
If we are here, there was no fatal error (but error may be still
|
|
|
|
unitialized).
|
|
|
|
*/
|
|
|
|
if (unlikely(entry->file->implicit_emptied))
|
|
|
|
{
|
|
|
|
entry->file->implicit_emptied= 0;
|
|
|
|
if (mysql_bin_log.is_open())
|
|
|
|
{
|
|
|
|
char *query, *end;
|
2005-11-23 21:45:02 +01:00
|
|
|
uint query_buf_size= 20 + share->db.length + share->table_name.length +1;
|
|
|
|
if ((query= (char*) my_malloc(query_buf_size,MYF(MY_WME))))
|
2004-06-25 15:52:01 +02:00
|
|
|
{
|
2005-12-22 06:39:02 +01:00
|
|
|
/* this DELETE FROM is needed even with row-based binlogging */
|
2004-06-25 15:52:01 +02:00
|
|
|
end = strxmov(strmov(query, "DELETE FROM `"),
|
2005-11-23 21:45:02 +01:00
|
|
|
share->db.str,"`.`",share->table_name.str,"`", NullS);
|
2009-05-30 15:32:28 +02:00
|
|
|
int errcode= query_error_code(thd, TRUE);
|
2005-12-22 06:39:02 +01:00
|
|
|
thd->binlog_query(THD::STMT_QUERY_TYPE,
|
2009-04-09 01:42:51 +02:00
|
|
|
query, (ulong)(end-query),
|
2009-05-30 15:32:28 +02:00
|
|
|
FALSE, FALSE, errcode);
|
2004-06-25 15:52:01 +02:00
|
|
|
my_free(query, MYF(0));
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
As replication is maybe going to be corrupted, we need to warn the
|
|
|
|
DBA on top of warning the client (which will automatically be done
|
|
|
|
because of MYF(MY_WME) in my_malloc() above).
|
|
|
|
*/
|
2005-11-23 21:45:02 +01:00
|
|
|
sql_print_error("When opening HEAP table, could not allocate memory "
|
|
|
|
"to write 'DELETE FROM `%s`.`%s`' to the binary log",
|
2009-11-30 16:55:03 +01:00
|
|
|
share->db.str, share->table_name.str);
|
2004-09-07 14:29:46 +02:00
|
|
|
delete entry->triggers;
|
2009-11-30 16:55:03 +01:00
|
|
|
return TRUE;
|
2004-06-25 15:52:01 +02:00
|
|
|
}
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Auxiliary routine which is used for performing automatical table repair.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static bool auto_repair_table(THD *thd, TABLE_LIST *table_list)
|
|
|
|
{
|
|
|
|
char cache_key[MAX_DBKEY_LENGTH];
|
|
|
|
uint cache_key_length;
|
|
|
|
TABLE_SHARE *share;
|
|
|
|
TABLE *entry;
|
|
|
|
int not_used;
|
|
|
|
bool result= FALSE;
|
|
|
|
|
|
|
|
cache_key_length= create_table_def_key(thd, cache_key, table_list, 0);
|
|
|
|
|
|
|
|
thd->clear_error();
|
|
|
|
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
|
|
|
|
if (!(share= get_table_share_with_create(thd, table_list, cache_key,
|
|
|
|
cache_key_length,
|
|
|
|
OPEN_VIEW, ¬_used)))
|
|
|
|
{
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (share->is_view)
|
|
|
|
goto end_with_lock_open;
|
|
|
|
|
|
|
|
if (!(entry= (TABLE*)my_malloc(sizeof(TABLE), MYF(MY_WME))))
|
|
|
|
{
|
|
|
|
result= TRUE;
|
|
|
|
goto end_with_lock_open;
|
|
|
|
}
|
|
|
|
share->version= 0;
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
2005-11-23 21:45:02 +01:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
if (open_table_from_share(thd, share, table_list->alias,
|
|
|
|
(uint) (HA_OPEN_KEYFILE | HA_OPEN_RNDFILE |
|
|
|
|
HA_GET_INDEX |
|
|
|
|
HA_TRY_READ_ONLY),
|
|
|
|
READ_KEYINFO | COMPUTE_TYPES | EXTRA_RECORD,
|
|
|
|
ha_open_options | HA_OPEN_FOR_REPAIR,
|
|
|
|
entry, FALSE) || ! entry->file ||
|
|
|
|
(entry->file->is_crashed() && entry->file->ha_check_and_repair(thd)))
|
|
|
|
{
|
|
|
|
/* Give right error message */
|
|
|
|
thd->clear_error();
|
|
|
|
my_error(ER_NOT_KEYFILE, MYF(0), share->table_name.str, my_errno);
|
|
|
|
sql_print_error("Couldn't repair table: %s.%s", share->db.str,
|
|
|
|
share->table_name.str);
|
|
|
|
if (entry->file)
|
|
|
|
closefrm(entry, 0);
|
|
|
|
result= TRUE;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
thd->clear_error(); // Clear error message
|
|
|
|
closefrm(entry, 0);
|
|
|
|
}
|
|
|
|
my_free(entry, MYF(0));
|
|
|
|
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
|
|
|
|
end_with_lock_open:
|
2009-11-30 20:11:32 +01:00
|
|
|
release_table_share(share);
|
2009-11-30 16:55:03 +01:00
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
/** Open_table_context */
|
|
|
|
|
|
|
|
Open_table_context::Open_table_context(THD *thd)
|
|
|
|
:m_action(OT_NO_ACTION),
|
|
|
|
m_can_deadlock(thd->in_multi_stmt_transaction() &&
|
|
|
|
thd->mdl_context.has_locks())
|
|
|
|
{}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Check if we can back-off and set back off action if we can.
|
|
|
|
Otherwise report and return error.
|
|
|
|
|
|
|
|
@retval TRUE if back-off is impossible.
|
|
|
|
@retval FALSE if we can back off. Back off action has been set.
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool
|
|
|
|
Open_table_context::
|
|
|
|
request_backoff_action(enum_open_table_action action_arg)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
We have met a exclusive metadata lock or a old version of
|
|
|
|
table and we are inside a transaction that already hold locks.
|
|
|
|
We can't follow the locking protocol in this scenario as it
|
|
|
|
might lead to deadlocks.
|
|
|
|
*/
|
|
|
|
if (m_can_deadlock)
|
|
|
|
{
|
|
|
|
my_error(ER_LOCK_DEADLOCK, MYF(0));
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
m_action= action_arg;
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/**
|
2009-12-01 16:20:43 +01:00
|
|
|
Recover from failed attempt ot open table by performing requested action.
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
@param thd Thread context
|
|
|
|
@param table Table list element for table that caused problem
|
|
|
|
@param action Type of action requested by failed open_table() call
|
|
|
|
|
2009-12-01 16:20:43 +01:00
|
|
|
@pre This function should be called only with "action" != OT_NO_ACTION.
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
@retval FALSE - Success. One should try to open tables once again.
|
|
|
|
@retval TRUE - Error
|
|
|
|
*/
|
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
bool
|
|
|
|
Open_table_context::
|
|
|
|
recover_from_failed_open_table_attempt(THD *thd, TABLE_LIST *table)
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
|
|
|
bool result= FALSE;
|
2009-12-08 10:57:07 +01:00
|
|
|
/* Execute the action. */
|
|
|
|
switch (m_action)
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
2009-12-08 10:57:07 +01:00
|
|
|
case OT_WAIT:
|
|
|
|
result= (thd->mdl_context.wait_for_locks(&m_mdl_requests) ||
|
|
|
|
tdc_wait_for_old_versions(thd, &m_mdl_requests));
|
2009-11-30 16:55:03 +01:00
|
|
|
break;
|
|
|
|
case OT_DISCOVER:
|
2009-12-02 17:15:40 +01:00
|
|
|
{
|
2009-12-08 10:57:07 +01:00
|
|
|
MDL_request mdl_xlock_request(&table->mdl_request);
|
|
|
|
mdl_xlock_request.set_type(MDL_EXCLUSIVE);
|
|
|
|
if ((result=
|
|
|
|
thd->mdl_context.acquire_exclusive_lock(&mdl_xlock_request)))
|
|
|
|
break;
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
tdc_remove_table(thd, TDC_RT_REMOVE_ALL, table->db, table->table_name);
|
|
|
|
ha_create_table_from_engine(thd, table->db, table->table_name);
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
|
|
|
|
thd->warning_info->clear_warning_info(thd->query_id);
|
|
|
|
thd->clear_error(); // Clear error message
|
|
|
|
thd->mdl_context.release_lock(mdl_xlock_request.ticket);
|
|
|
|
break;
|
2009-12-02 17:15:40 +01:00
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
case OT_REPAIR:
|
2009-12-02 17:15:40 +01:00
|
|
|
{
|
2009-12-08 10:57:07 +01:00
|
|
|
MDL_request mdl_xlock_request(&table->mdl_request);
|
|
|
|
mdl_xlock_request.set_type(MDL_EXCLUSIVE);
|
|
|
|
if ((result=
|
|
|
|
thd->mdl_context.acquire_exclusive_lock(&mdl_xlock_request)))
|
|
|
|
break;
|
2009-11-30 16:55:03 +01:00
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
tdc_remove_table(thd, TDC_RT_REMOVE_ALL, table->db, table->table_name);
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
|
|
|
|
result= auto_repair_table(thd, table);
|
|
|
|
thd->mdl_context.release_lock(mdl_xlock_request.ticket);
|
|
|
|
break;
|
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
default:
|
|
|
|
DBUG_ASSERT(0);
|
|
|
|
}
|
2009-12-08 10:57:07 +01:00
|
|
|
/* Remove all old requests, they will be re-added. */
|
|
|
|
m_mdl_requests.empty();
|
|
|
|
/* Prepare for possible another back-off. */
|
|
|
|
m_action= OT_NO_ACTION;
|
2009-11-30 16:55:03 +01:00
|
|
|
return result;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2005-07-30 10:19:57 +02:00
|
|
|
|
2008-09-29 15:53:40 +02:00
|
|
|
/*
|
|
|
|
Return a appropriate read lock type given a table object.
|
|
|
|
|
|
|
|
@param thd Thread context
|
|
|
|
@param table TABLE object for table to be locked
|
|
|
|
|
|
|
|
@remark Due to a statement-based replication limitation, statements such as
|
|
|
|
INSERT INTO .. SELECT FROM .. and CREATE TABLE .. SELECT FROM need
|
|
|
|
to grab a TL_READ_NO_INSERT lock on the source table in order to
|
|
|
|
prevent the replication of a concurrent statement that modifies the
|
|
|
|
source table. If such a statement gets applied on the slave before
|
|
|
|
the INSERT .. SELECT statement finishes, data on the master could
|
|
|
|
differ from data on the slave and end-up with a discrepancy between
|
|
|
|
the binary log and table state. Furthermore, this does not apply to
|
|
|
|
I_S and log tables as it's always unsafe to replicate such tables
|
|
|
|
under statement-based replication as the table on the slave might
|
|
|
|
contain other data (ie: general_log is enabled on the slave). The
|
|
|
|
statement will be marked as unsafe for SBR in decide_logging_format().
|
|
|
|
*/
|
|
|
|
|
|
|
|
thr_lock_type read_lock_type_for_table(THD *thd, TABLE *table)
|
|
|
|
{
|
|
|
|
bool log_on= mysql_bin_log.is_open() && (thd->options & OPTION_BIN_LOG);
|
|
|
|
ulong binlog_format= thd->variables.binlog_format;
|
|
|
|
if ((log_on == FALSE) || (binlog_format == BINLOG_FORMAT_ROW) ||
|
|
|
|
(table->s->table_category == TABLE_CATEGORY_PERFORMANCE))
|
|
|
|
return TL_READ;
|
|
|
|
else
|
|
|
|
return TL_READ_NO_INSERT;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2004-06-25 15:52:01 +02:00
|
|
|
/*
|
2009-12-09 17:00:46 +01:00
|
|
|
Handle element of prelocking set other than table. E.g. cache routine
|
|
|
|
and, if prelocking strategy prescribes so, extend the prelocking set
|
|
|
|
with tables and routines used by it.
|
2009-12-08 15:13:12 +01:00
|
|
|
|
|
|
|
@param[in] thd Thread context.
|
|
|
|
@param[in] prelocking_ctx Prelocking context.
|
2009-12-09 17:00:46 +01:00
|
|
|
@param[in] rt Element of prelocking set to be processed.
|
2009-12-08 15:13:12 +01:00
|
|
|
@param[in] prelocking_strategy Strategy which specifies how the
|
|
|
|
prelocking set should be extended when
|
|
|
|
one of its elements is processed.
|
|
|
|
@param[out] need_prelocking Set to TRUE if it was detected that this
|
|
|
|
statement will require prelocked mode for
|
|
|
|
its execution, not touched otherwise.
|
|
|
|
|
|
|
|
@retval FALSE Success.
|
|
|
|
@retval TRUE Failure (Conflicting metadata lock, OOM, other errors).
|
|
|
|
*/
|
|
|
|
|
|
|
|
static bool
|
2009-12-09 17:00:46 +01:00
|
|
|
open_and_process_routine(THD *thd, Query_tables_list *prelocking_ctx,
|
|
|
|
Sroutine_hash_entry *rt,
|
|
|
|
Prelocking_strategy *prelocking_strategy,
|
|
|
|
bool *need_prelocking)
|
2009-12-08 15:13:12 +01:00
|
|
|
{
|
2009-12-09 17:00:46 +01:00
|
|
|
DBUG_ENTER("open_and_process_routine");
|
2009-12-08 15:13:12 +01:00
|
|
|
|
2009-12-09 17:11:26 +01:00
|
|
|
switch (rt->mdl_request.key.mdl_namespace())
|
2009-12-09 17:00:46 +01:00
|
|
|
{
|
2009-12-10 09:21:38 +01:00
|
|
|
case MDL_key::FUNCTION:
|
|
|
|
case MDL_key::PROCEDURE:
|
2009-12-08 15:13:12 +01:00
|
|
|
{
|
2009-12-09 17:11:26 +01:00
|
|
|
char qname_buff[NAME_LEN*2+1+1];
|
|
|
|
sp_name name(&rt->mdl_request.key, qname_buff);
|
2009-12-09 17:00:46 +01:00
|
|
|
sp_head *sp;
|
2009-12-10 09:21:38 +01:00
|
|
|
int type= (rt->mdl_request.key.mdl_namespace() == MDL_key::FUNCTION) ?
|
2009-12-09 17:11:26 +01:00
|
|
|
TYPE_ENUM_FUNCTION : TYPE_ENUM_PROCEDURE;
|
2009-12-08 15:13:12 +01:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
if (sp_cache_routine(thd, type, &name, &sp))
|
|
|
|
DBUG_RETURN(TRUE);
|
2009-12-08 15:13:12 +01:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
if (sp)
|
|
|
|
{
|
|
|
|
prelocking_strategy->handle_routine(thd, prelocking_ctx, rt, sp,
|
|
|
|
need_prelocking);
|
2009-12-08 15:13:12 +01:00
|
|
|
}
|
|
|
|
}
|
2009-12-09 17:00:46 +01:00
|
|
|
break;
|
2009-12-10 09:21:38 +01:00
|
|
|
case MDL_key::TRIGGER:
|
2009-12-09 17:00:46 +01:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
/* Impossible type value. */
|
|
|
|
DBUG_ASSERT(0);
|
2009-12-08 15:13:12 +01:00
|
|
|
}
|
|
|
|
DBUG_RETURN(FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
/**
|
|
|
|
Handle table list element by obtaining metadata lock, opening table or view
|
|
|
|
and, if prelocking strategy prescribes so, extending the prelocking set with
|
|
|
|
tables and routines used by it.
|
|
|
|
|
|
|
|
@param[in] thd Thread context.
|
|
|
|
@param[in] lex LEX structure for statement.
|
|
|
|
@param[in] tables Table list element to be processed.
|
|
|
|
@param[in,out] counter Number of tables which are open.
|
|
|
|
@param[in] flags Bitmap of flags to modify how the tables
|
|
|
|
will be open, see open_table() description
|
|
|
|
for details.
|
|
|
|
@param[in] prelocking_strategy Strategy which specifies how the
|
|
|
|
prelocking set should be extended
|
|
|
|
when table or view is processed.
|
|
|
|
@param[in] has_prelocking_list Indicates that prelocking set/list for
|
|
|
|
this statement has already been built.
|
|
|
|
@param[in] ot_ctx Context used to recover from a failed
|
|
|
|
open_table() attempt.
|
|
|
|
@param[in] new_frm_mem Temporary MEM_ROOT to be used for
|
|
|
|
parsing .FRMs for views.
|
|
|
|
|
|
|
|
@retval FALSE Success.
|
|
|
|
@retval TRUE Error, reported unless there is a chance to recover from it.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static bool
|
|
|
|
open_and_process_table(THD *thd, LEX *lex, TABLE_LIST *tables,
|
|
|
|
uint *counter, uint flags,
|
|
|
|
Prelocking_strategy *prelocking_strategy,
|
|
|
|
bool has_prelocking_list,
|
|
|
|
Open_table_context *ot_ctx,
|
|
|
|
MEM_ROOT *new_frm_mem)
|
|
|
|
{
|
|
|
|
bool error= FALSE;
|
|
|
|
bool safe_to_ignore_table= FALSE;
|
|
|
|
DBUG_ENTER("open_and_process_table");
|
|
|
|
|
|
|
|
/*
|
|
|
|
Ignore placeholders for derived tables. After derived tables
|
|
|
|
processing, link to created temporary table will be put here.
|
|
|
|
If this is derived table for view then we still want to process
|
|
|
|
routines used by this view.
|
|
|
|
*/
|
|
|
|
if (tables->derived)
|
|
|
|
{
|
|
|
|
if (!tables->view)
|
|
|
|
goto end;
|
|
|
|
/*
|
|
|
|
We restore view's name and database wiped out by derived tables
|
|
|
|
processing and fall back to standard open process in order to
|
|
|
|
obtain proper metadata locks and do other necessary steps like
|
|
|
|
stored routine processing.
|
|
|
|
*/
|
|
|
|
tables->db= tables->view_db.str;
|
|
|
|
tables->db_length= tables->view_db.length;
|
|
|
|
tables->table_name= tables->view_name.str;
|
|
|
|
tables->table_name_length= tables->view_name.length;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
If this TABLE_LIST object is a placeholder for an information_schema
|
|
|
|
table, create a temporary table to represent the information_schema
|
|
|
|
table in the query. Do not fill it yet - will be filled during
|
|
|
|
execution.
|
|
|
|
*/
|
|
|
|
if (tables->schema_table)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
If this information_schema table is merged into a mergeable
|
|
|
|
view, ignore it for now -- it will be filled when its respective
|
|
|
|
TABLE_LIST is processed. This code works only during re-execution.
|
|
|
|
*/
|
|
|
|
if (tables->view)
|
|
|
|
goto process_view_routines;
|
|
|
|
if (!mysql_schema_table(thd, lex, tables) &&
|
|
|
|
!check_and_update_table_version(thd, tables, tables->table->s))
|
|
|
|
{
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
error= TRUE;
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
DBUG_PRINT("tcache", ("opening table: '%s'.'%s' item: %p",
|
|
|
|
tables->db, tables->table_name, tables)); //psergey: invalid read of size 1 here
|
|
|
|
(*counter)++;
|
|
|
|
|
|
|
|
/* Not a placeholder: must be a base table or a view. Let us open it. */
|
|
|
|
DBUG_ASSERT(!tables->table);
|
|
|
|
|
|
|
|
if (tables->prelocking_placeholder)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
For the tables added by the pre-locking code, attempt to open
|
|
|
|
the table but fail silently if the table does not exist.
|
|
|
|
The real failure will occur when/if a statement attempts to use
|
|
|
|
that table.
|
|
|
|
*/
|
|
|
|
Prelock_error_handler prelock_handler;
|
|
|
|
thd->push_internal_handler(& prelock_handler);
|
|
|
|
error= open_table(thd, tables, new_frm_mem, ot_ctx, flags);
|
|
|
|
thd->pop_internal_handler();
|
|
|
|
safe_to_ignore_table= prelock_handler.safely_trapped_errors();
|
|
|
|
}
|
|
|
|
else
|
|
|
|
error= open_table(thd, tables, new_frm_mem, ot_ctx, flags);
|
|
|
|
|
|
|
|
free_root(new_frm_mem, MYF(MY_KEEP_PREALLOC));
|
|
|
|
|
|
|
|
if (error)
|
|
|
|
{
|
|
|
|
if (! ot_ctx->can_recover_from_failed_open_table() && safe_to_ignore_table)
|
|
|
|
{
|
|
|
|
DBUG_PRINT("info", ("open_table: ignoring table '%s'.'%s'",
|
|
|
|
tables->db, tables->alias));
|
|
|
|
error= FALSE;
|
|
|
|
}
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
We can't rely on simple check for TABLE_LIST::view to determine
|
|
|
|
that this is a view since during re-execution we might reopen
|
|
|
|
ordinary table in place of view and thus have TABLE_LIST::view
|
|
|
|
set from repvious execution and TABLE_LIST::table set from
|
|
|
|
current.
|
|
|
|
*/
|
|
|
|
if (!tables->table && tables->view)
|
|
|
|
{
|
|
|
|
/* VIEW placeholder */
|
|
|
|
(*counter)--;
|
|
|
|
|
|
|
|
/*
|
|
|
|
tables->next_global list consists of two parts:
|
|
|
|
1) Query tables and underlying tables of views.
|
|
|
|
2) Tables used by all stored routines that this statement invokes on
|
|
|
|
execution.
|
|
|
|
We need to know where the bound between these two parts is. If we've
|
|
|
|
just opened a view, which was the last table in part #1, and it
|
|
|
|
has added its base tables after itself, adjust the boundary pointer
|
|
|
|
accordingly.
|
|
|
|
*/
|
|
|
|
if (lex->query_tables_own_last == &(tables->next_global) &&
|
|
|
|
tables->view->query_tables)
|
|
|
|
lex->query_tables_own_last= tables->view->query_tables_last;
|
|
|
|
/*
|
|
|
|
Let us free memory used by 'sroutines' hash here since we never
|
|
|
|
call destructor for this LEX.
|
|
|
|
*/
|
|
|
|
my_hash_free(&tables->view->sroutines);
|
|
|
|
goto process_view_routines;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
Special types of open can succeed but still don't set
|
|
|
|
TABLE_LIST::table to anything.
|
|
|
|
*/
|
|
|
|
if (tables->open_strategy && !tables->table)
|
|
|
|
goto end;
|
|
|
|
|
|
|
|
/*
|
|
|
|
If we are not already in prelocked mode and extended table list is not
|
|
|
|
yet built we might have to build the prelocking set for this statement.
|
|
|
|
|
|
|
|
Since currently no prelocking strategy prescribes doing anything for
|
|
|
|
tables which are only read, we do below checks only if table is going
|
|
|
|
to be changed.
|
|
|
|
*/
|
|
|
|
if (thd->locked_tables_mode <= LTM_LOCK_TABLES &&
|
|
|
|
! has_prelocking_list &&
|
|
|
|
tables->lock_type >= TL_WRITE_ALLOW_WRITE)
|
|
|
|
{
|
|
|
|
bool need_prelocking= FALSE;
|
|
|
|
TABLE_LIST **save_query_tables_last= lex->query_tables_last;
|
|
|
|
/*
|
|
|
|
Extend statement's table list and the prelocking set with
|
|
|
|
tables and routines according to the current prelocking
|
|
|
|
strategy.
|
|
|
|
|
|
|
|
For example, for DML statements we need to add tables and routines
|
|
|
|
used by triggers which are going to be invoked for this element of
|
|
|
|
table list and also add tables required for handling of foreign keys.
|
|
|
|
*/
|
|
|
|
error= prelocking_strategy->handle_table(thd, lex, tables,
|
|
|
|
&need_prelocking);
|
|
|
|
|
|
|
|
if (need_prelocking && ! lex->requires_prelocking())
|
|
|
|
lex->mark_as_requiring_prelocking(save_query_tables_last);
|
|
|
|
|
|
|
|
if (error)
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (tables->lock_type != TL_UNLOCK && ! thd->locked_tables_mode)
|
|
|
|
{
|
|
|
|
if (tables->lock_type == TL_WRITE_DEFAULT)
|
|
|
|
tables->table->reginfo.lock_type= thd->update_lock_default;
|
|
|
|
else if (tables->lock_type == TL_READ_DEFAULT)
|
|
|
|
tables->table->reginfo.lock_type=
|
|
|
|
read_lock_type_for_table(thd, tables->table);
|
|
|
|
else
|
|
|
|
tables->table->reginfo.lock_type= tables->lock_type;
|
|
|
|
}
|
|
|
|
tables->table->grant= tables->grant;
|
|
|
|
|
|
|
|
/* Check and update metadata version of a base table. */
|
|
|
|
error= check_and_update_table_version(thd, tables, tables->table->s);
|
|
|
|
|
|
|
|
if (error)
|
|
|
|
goto end;
|
|
|
|
/*
|
|
|
|
After opening a MERGE table add the children to the query list of
|
|
|
|
tables, so that they are opened too.
|
|
|
|
Note that placeholders don't have the handler open.
|
|
|
|
*/
|
|
|
|
/* MERGE tables need to access parent and child TABLE_LISTs. */
|
|
|
|
DBUG_ASSERT(tables->table->pos_in_table_list == tables);
|
|
|
|
/* Non-MERGE tables ignore this call. */
|
|
|
|
if (tables->table->file->extra(HA_EXTRA_ADD_CHILDREN_LIST))
|
|
|
|
{
|
|
|
|
error= TRUE;
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
|
|
|
|
process_view_routines:
|
|
|
|
/*
|
|
|
|
Again we may need cache all routines used by this view and add
|
|
|
|
tables used by them to table list.
|
|
|
|
*/
|
|
|
|
if (tables->view &&
|
|
|
|
thd->locked_tables_mode <= LTM_LOCK_TABLES &&
|
|
|
|
! has_prelocking_list)
|
|
|
|
{
|
|
|
|
bool need_prelocking= FALSE;
|
|
|
|
TABLE_LIST **save_query_tables_last= lex->query_tables_last;
|
|
|
|
|
|
|
|
error= prelocking_strategy->handle_view(thd, lex, tables,
|
|
|
|
&need_prelocking);
|
|
|
|
|
|
|
|
if (need_prelocking && ! lex->requires_prelocking())
|
|
|
|
lex->mark_as_requiring_prelocking(save_query_tables_last);
|
|
|
|
|
|
|
|
if (error)
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
|
|
|
|
end:
|
|
|
|
DBUG_RETURN(error);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-08 15:13:12 +01:00
|
|
|
/**
|
2004-06-25 15:52:01 +02:00
|
|
|
Open all tables in list
|
|
|
|
|
2009-12-08 15:13:12 +01:00
|
|
|
@param[in] thd Thread context.
|
|
|
|
@param[in,out] start List of tables to be open (it can be adjusted for
|
|
|
|
statement that uses tables only implicitly, e.g.
|
|
|
|
for "SELECT f1()").
|
|
|
|
@param[out] counter Number of tables which were open.
|
|
|
|
@param[in] flags Bitmap of flags to modify how the tables will be
|
|
|
|
open, see open_table() description for details.
|
|
|
|
@param[in] prelocking_strategy Strategy which specifies how prelocking
|
|
|
|
algorithm should work for this statement.
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2009-12-08 15:13:12 +01:00
|
|
|
@note
|
|
|
|
Unless we are already in prelocked mode and prelocking strategy prescribes
|
|
|
|
so this function will also precache all SP/SFs explicitly or implicitly
|
|
|
|
(via views and triggers) used by the query and add tables needed for their
|
|
|
|
execution to table list. Statement that uses SFs, invokes triggers or
|
|
|
|
requires foreign key checks will be marked as requiring prelocking.
|
2005-03-04 14:35:28 +01:00
|
|
|
Prelocked mode will be enabled for such query during lock_tables() call.
|
|
|
|
|
|
|
|
If query for which we are opening tables is already marked as requiring
|
|
|
|
prelocking it won't do such precaching and will simply reuse table list
|
|
|
|
which is already built.
|
|
|
|
|
2009-12-09 10:04:55 +01:00
|
|
|
@retval FALSE Success.
|
|
|
|
@retval TRUE Error, reported.
|
2004-06-25 15:52:01 +02:00
|
|
|
*/
|
|
|
|
|
2009-12-09 10:04:55 +01:00
|
|
|
bool open_tables(THD *thd, TABLE_LIST **start, uint *counter, uint flags,
|
2009-12-08 15:13:12 +01:00
|
|
|
Prelocking_strategy *prelocking_strategy)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-12-09 17:00:46 +01:00
|
|
|
/*
|
|
|
|
We use pointers to "next_global" member in the last processed TABLE_LIST
|
|
|
|
element and to the "next" member in the last processed Sroutine_hash_entry
|
|
|
|
element as iterators over, correspondingly, the table list and stored routines
|
|
|
|
list which stay valid and allow to continue iteration when new elements are
|
|
|
|
added to the tail of the lists.
|
|
|
|
*/
|
|
|
|
TABLE_LIST **table_to_open;
|
|
|
|
Sroutine_hash_entry **sroutine_to_open;
|
|
|
|
TABLE_LIST *tables;
|
2009-12-08 10:57:07 +01:00
|
|
|
Open_table_context ot_ctx(thd);
|
2009-12-09 10:04:55 +01:00
|
|
|
bool error= FALSE;
|
2004-07-16 00:15:55 +02:00
|
|
|
MEM_ROOT new_frm_mem;
|
2009-12-09 17:00:46 +01:00
|
|
|
bool has_prelocking_list= thd->lex->requires_prelocking();
|
2005-03-04 14:35:28 +01:00
|
|
|
DBUG_ENTER("open_tables");
|
2009-11-30 16:55:03 +01:00
|
|
|
/*
|
|
|
|
Close HANDLER tables which are marked for flush or against which there
|
|
|
|
are pending exclusive metadata locks. Note that we do this not to avoid
|
|
|
|
deadlocks (calls to mysql_ha_flush() in mdl_wait_for_locks() and
|
|
|
|
tdc_wait_for_old_version() are enough for this) but in order to have
|
|
|
|
a point during statement execution at which such HANDLERs are closed
|
|
|
|
even if they don't create problems for current thread (i.e. to avoid
|
|
|
|
having DDL blocked by HANDLERs opened for long time).
|
|
|
|
*/
|
|
|
|
if (thd->handler_tables)
|
|
|
|
mysql_ha_flush(thd);
|
|
|
|
|
2005-03-04 14:35:28 +01:00
|
|
|
/*
|
2009-12-09 17:00:46 +01:00
|
|
|
temporary mem_root for new .frm parsing.
|
|
|
|
TODO: variables for size
|
2005-03-04 14:35:28 +01:00
|
|
|
*/
|
2009-12-09 17:00:46 +01:00
|
|
|
init_sql_alloc(&new_frm_mem, 8024, 8024);
|
2005-10-27 23:18:23 +02:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
thd->current_tablenr= 0;
|
|
|
|
restart:
|
|
|
|
table_to_open= start;
|
|
|
|
sroutine_to_open= (Sroutine_hash_entry**) &thd->lex->sroutines_list.first;
|
|
|
|
*counter= 0;
|
|
|
|
thd_proc_info(thd, "Opening tables");
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2007-02-28 14:03:38 +01:00
|
|
|
/*
|
2009-12-09 17:00:46 +01:00
|
|
|
Perform steps of prelocking algorithm until there are unprocessed
|
|
|
|
elements in prelocking list/set.
|
2007-02-28 14:03:38 +01:00
|
|
|
*/
|
2009-12-09 17:00:46 +01:00
|
|
|
while (*table_to_open ||
|
|
|
|
(thd->locked_tables_mode <= LTM_LOCK_TABLES &&
|
|
|
|
! has_prelocking_list &&
|
|
|
|
*sroutine_to_open))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2007-02-28 14:03:38 +01:00
|
|
|
/*
|
2009-12-09 17:00:46 +01:00
|
|
|
For every table in the list of tables to open, try to find or open
|
|
|
|
a table.
|
2007-02-28 14:03:38 +01:00
|
|
|
*/
|
2009-12-09 17:00:46 +01:00
|
|
|
for (tables= *table_to_open; tables;
|
|
|
|
table_to_open= &tables->next_global, tables= tables->next_global)
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
{
|
2009-12-09 17:00:46 +01:00
|
|
|
error= open_and_process_table(thd, thd->lex, tables, counter,
|
|
|
|
flags, prelocking_strategy,
|
|
|
|
has_prelocking_list, &ot_ctx,
|
|
|
|
&new_frm_mem);
|
2004-07-16 00:15:55 +02:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
if (error)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-12-09 17:00:46 +01:00
|
|
|
if (ot_ctx.can_recover_from_failed_open_table())
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
We have met exclusive metadata lock or old version of table.
|
|
|
|
Now we have to close all tables and release metadata locks.
|
|
|
|
We also have to throw away set of prelocked tables (and thus
|
|
|
|
close tables from this set that were open by now) since it
|
|
|
|
is possible that one of tables which determined its content
|
|
|
|
was changed.
|
|
|
|
|
|
|
|
Instead of implementing complex/non-robust logic mentioned
|
|
|
|
above we simply close and then reopen all tables.
|
|
|
|
|
|
|
|
We have to save pointer to table list element for table which we
|
|
|
|
have failed to open since closing tables can trigger removal of
|
|
|
|
elements from the table list (if MERGE tables are involved),
|
|
|
|
*/
|
|
|
|
TABLE_LIST *failed_table= *table_to_open;
|
|
|
|
close_tables_for_reopen(thd, start);
|
|
|
|
/*
|
|
|
|
Here we rely on the fact that 'tables' still points to the valid
|
|
|
|
TABLE_LIST element. Altough currently this assumption is valid
|
|
|
|
it may change in future.
|
|
|
|
*/
|
|
|
|
if (ot_ctx.recover_from_failed_open_table_attempt(thd, failed_table))
|
|
|
|
goto err;
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
error= FALSE;
|
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
goto err;
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2009-12-01 15:58:31 +01:00
|
|
|
|
|
|
|
/*
|
2009-12-09 17:00:46 +01:00
|
|
|
If we are not already in prelocked mode and extended table list is
|
|
|
|
not yet built for our statement we need to cache routines it uses
|
|
|
|
and build the prelocking list for it.
|
2009-12-01 15:58:31 +01:00
|
|
|
*/
|
2009-12-09 17:00:46 +01:00
|
|
|
if (thd->locked_tables_mode <= LTM_LOCK_TABLES && ! has_prelocking_list)
|
2009-12-01 15:58:31 +01:00
|
|
|
{
|
2009-12-08 15:13:12 +01:00
|
|
|
bool need_prelocking= FALSE;
|
|
|
|
TABLE_LIST **save_query_tables_last= thd->lex->query_tables_last;
|
|
|
|
/*
|
2009-12-09 17:00:46 +01:00
|
|
|
Process elements of the prelocking set which are present there
|
|
|
|
since parsing stage or were added to it by invocations of
|
|
|
|
Prelocking_strategy methods in the above loop over tables.
|
2009-12-08 15:13:12 +01:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
For example, if element is a routine, cache it and then,
|
|
|
|
if prelocking strategy prescribes so, add tables it uses to the
|
|
|
|
table list and routines it might invoke to the prelocking set.
|
2009-12-08 15:13:12 +01:00
|
|
|
*/
|
2009-12-09 17:00:46 +01:00
|
|
|
for (Sroutine_hash_entry *rt= *sroutine_to_open; rt;
|
|
|
|
sroutine_to_open= &rt->next, rt= rt->next)
|
|
|
|
{
|
|
|
|
error= open_and_process_routine(thd, thd->lex, rt,
|
|
|
|
prelocking_strategy,
|
|
|
|
&need_prelocking);
|
2009-12-08 15:13:12 +01:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
if (error)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
Serious error during reading stored routines from mysql.proc table.
|
|
|
|
Something is wrong with the table or its contents, and an error has
|
|
|
|
been emitted; we must abort.
|
|
|
|
*/
|
|
|
|
goto err;
|
|
|
|
}
|
|
|
|
}
|
2009-12-09 10:04:55 +01:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
if (need_prelocking && ! thd->lex->requires_prelocking())
|
|
|
|
thd->lex->mark_as_requiring_prelocking(save_query_tables_last);
|
2009-12-09 10:04:55 +01:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
if (need_prelocking && ! *start)
|
|
|
|
*start= thd->lex->query_tables;
|
2005-09-15 01:56:09 +02:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2005-10-26 15:34:57 +02:00
|
|
|
|
2009-12-03 00:09:22 +01:00
|
|
|
/*
|
|
|
|
After successful open of all tables, including MERGE parents and
|
|
|
|
children, attach the children to their parents. At end of statement,
|
|
|
|
the children are detached. Attaching and detaching are always done,
|
|
|
|
even under LOCK TABLES.
|
|
|
|
*/
|
|
|
|
for (tables= *start; tables; tables= tables->next_global)
|
|
|
|
{
|
|
|
|
TABLE *tbl= tables->table;
|
|
|
|
|
|
|
|
/* Schema tables may not have a TABLE object here. */
|
|
|
|
if (tbl && tbl->file->ht->db_type == DB_TYPE_MRG_MYISAM)
|
|
|
|
{
|
|
|
|
/* MERGE tables need to access parent and child TABLE_LISTs. */
|
|
|
|
DBUG_ASSERT(tbl->pos_in_table_list == tables);
|
|
|
|
if (tbl->file->extra(HA_EXTRA_ATTACH_CHILDREN))
|
|
|
|
{
|
2009-12-09 10:04:55 +01:00
|
|
|
error= TRUE;
|
2009-12-03 00:09:22 +01:00
|
|
|
goto err;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2009-12-09 10:04:55 +01:00
|
|
|
err:
|
2007-02-22 16:03:08 +01:00
|
|
|
thd_proc_info(thd, 0);
|
2004-09-03 20:43:04 +02:00
|
|
|
free_root(&new_frm_mem, MYF(0)); // Free pre-alloced block
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2009-12-09 17:00:46 +01:00
|
|
|
if (error && *table_to_open)
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
{
|
2009-12-09 17:00:46 +01:00
|
|
|
(*table_to_open)->table= NULL;
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
}
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_PRINT("open_tables", ("returning: %d", (int) error));
|
|
|
|
DBUG_RETURN(error);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-12-08 15:13:12 +01:00
|
|
|
/**
|
|
|
|
Defines how prelocking algorithm for DML statements should handle routines:
|
|
|
|
- For CALL statements we do unrolling (i.e. open and lock tables for each
|
|
|
|
sub-statement individually). So for such statements prelocking is enabled
|
|
|
|
only if stored functions are used in parameter list and only for period
|
|
|
|
during which we calculate values of parameters. Thus in this strategy we
|
|
|
|
ignore procedure which is directly called by such statement and extend
|
|
|
|
the prelocking set only with tables/functions used by SF called from the
|
|
|
|
parameter list.
|
|
|
|
- For any other statement any routine which is directly or indirectly called
|
|
|
|
by statement is going to be executed in prelocked mode. So in this case we
|
|
|
|
simply add all tables and routines used by it to the prelocking set.
|
|
|
|
|
|
|
|
@param[in] thd Thread context.
|
|
|
|
@param[in] prelocking_ctx Prelocking context of the statement.
|
|
|
|
@param[in] rt Prelocking set element describing routine.
|
|
|
|
@param[in] sp Routine body.
|
|
|
|
@param[out] need_prelocking Set to TRUE if method detects that prelocking
|
|
|
|
required, not changed otherwise.
|
|
|
|
|
|
|
|
@retval FALSE Success.
|
|
|
|
@retval TRUE Failure (OOM).
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool DML_prelocking_strategy::
|
|
|
|
handle_routine(THD *thd, Query_tables_list *prelocking_ctx,
|
|
|
|
Sroutine_hash_entry *rt, sp_head *sp, bool *need_prelocking)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
We assume that for any "CALL proc(...)" statement sroutines_list will
|
|
|
|
have 'proc' as first element (it may have several, consider e.g.
|
|
|
|
"proc(sp_func(...)))". This property is currently guaranted by the
|
|
|
|
parser.
|
|
|
|
*/
|
|
|
|
|
|
|
|
if (rt != (Sroutine_hash_entry*)prelocking_ctx->sroutines_list.first ||
|
2009-12-10 09:21:38 +01:00
|
|
|
rt->mdl_request.key.mdl_namespace() != MDL_key::PROCEDURE)
|
2009-12-08 15:13:12 +01:00
|
|
|
{
|
|
|
|
*need_prelocking= TRUE;
|
|
|
|
sp_update_stmt_used_routines(thd, prelocking_ctx, &sp->m_sroutines,
|
|
|
|
rt->belong_to_view);
|
|
|
|
(void)sp->add_used_tables_to_table_list(thd,
|
|
|
|
&prelocking_ctx->query_tables_last,
|
|
|
|
rt->belong_to_view);
|
|
|
|
}
|
|
|
|
sp->propagate_attributes(prelocking_ctx);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Defines how prelocking algorithm for DML statements should handle table list
|
|
|
|
elements:
|
|
|
|
- If table has triggers we should add all tables and routines
|
|
|
|
used by them to the prelocking set.
|
|
|
|
|
|
|
|
We do not need to acquire metadata locks on trigger names
|
|
|
|
in DML statements, since all DDL statements
|
|
|
|
that change trigger metadata always lock their
|
|
|
|
subject tables.
|
|
|
|
|
|
|
|
@param[in] thd Thread context.
|
|
|
|
@param[in] prelocking_ctx Prelocking context of the statement.
|
|
|
|
@param[in] table_list Table list element for table.
|
|
|
|
@param[in] sp Routine body.
|
|
|
|
@param[out] need_prelocking Set to TRUE if method detects that prelocking
|
|
|
|
required, not changed otherwise.
|
|
|
|
|
|
|
|
@retval FALSE Success.
|
|
|
|
@retval TRUE Failure (OOM).
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool DML_prelocking_strategy::
|
|
|
|
handle_table(THD *thd, Query_tables_list *prelocking_ctx,
|
|
|
|
TABLE_LIST *table_list, bool *need_prelocking)
|
|
|
|
{
|
|
|
|
/* We rely on a caller to check that table is going to be changed. */
|
|
|
|
DBUG_ASSERT(table_list->lock_type >= TL_WRITE_ALLOW_WRITE);
|
|
|
|
|
|
|
|
if (table_list->trg_event_map)
|
|
|
|
{
|
|
|
|
if (table_list->table->triggers)
|
|
|
|
{
|
|
|
|
*need_prelocking= TRUE;
|
|
|
|
|
|
|
|
if (table_list->table->triggers->
|
|
|
|
add_tables_and_routines_for_triggers(thd, prelocking_ctx, table_list))
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Defines how prelocking algorithm for DML statements should handle view -
|
|
|
|
all view routines should be added to the prelocking set.
|
|
|
|
|
|
|
|
@param[in] thd Thread context.
|
|
|
|
@param[in] prelocking_ctx Prelocking context of the statement.
|
|
|
|
@param[in] table_list Table list element for view.
|
|
|
|
@param[in] sp Routine body.
|
|
|
|
@param[out] need_prelocking Set to TRUE if method detects that prelocking
|
|
|
|
required, not changed otherwise.
|
|
|
|
|
|
|
|
@retval FALSE Success.
|
|
|
|
@retval TRUE Failure (OOM).
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool DML_prelocking_strategy::
|
|
|
|
handle_view(THD *thd, Query_tables_list *prelocking_ctx,
|
|
|
|
TABLE_LIST *table_list, bool *need_prelocking)
|
|
|
|
{
|
|
|
|
if (table_list->view->uses_stored_routines())
|
|
|
|
{
|
|
|
|
*need_prelocking= TRUE;
|
|
|
|
|
|
|
|
sp_update_stmt_used_routines(thd, prelocking_ctx,
|
|
|
|
&table_list->view->sroutines_list,
|
|
|
|
table_list->top_table());
|
|
|
|
}
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Defines how prelocking algorithm for LOCK TABLES statement should handle
|
|
|
|
table list elements.
|
|
|
|
|
|
|
|
@param[in] thd Thread context.
|
|
|
|
@param[in] prelocking_ctx Prelocking context of the statement.
|
|
|
|
@param[in] table_list Table list element for table.
|
|
|
|
@param[in] sp Routine body.
|
|
|
|
@param[out] need_prelocking Set to TRUE if method detects that prelocking
|
|
|
|
required, not changed otherwise.
|
|
|
|
|
|
|
|
@retval FALSE Success.
|
|
|
|
@retval TRUE Failure (OOM).
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool Lock_tables_prelocking_strategy::
|
|
|
|
handle_table(THD *thd, Query_tables_list *prelocking_ctx,
|
|
|
|
TABLE_LIST *table_list, bool *need_prelocking)
|
|
|
|
{
|
|
|
|
if (DML_prelocking_strategy::handle_table(thd, prelocking_ctx, table_list,
|
|
|
|
need_prelocking))
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
/* We rely on a caller to check that table is going to be changed. */
|
|
|
|
DBUG_ASSERT(table_list->lock_type >= TL_WRITE_ALLOW_WRITE);
|
|
|
|
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Defines how prelocking algorithm for ALTER TABLE statement should handle
|
|
|
|
routines - do nothing as this statement is not supposed to call routines.
|
|
|
|
|
|
|
|
We still can end up in this method when someone tries
|
|
|
|
to define a foreign key referencing a view, and not just
|
|
|
|
a simple view, but one that uses stored routines.
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool Alter_table_prelocking_strategy::
|
|
|
|
handle_routine(THD *thd, Query_tables_list *prelocking_ctx,
|
|
|
|
Sroutine_hash_entry *rt, sp_head *sp, bool *need_prelocking)
|
|
|
|
{
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Defines how prelocking algorithm for ALTER TABLE statement should handle
|
|
|
|
table list elements.
|
|
|
|
|
|
|
|
Unlike in DML, we do not process triggers here.
|
|
|
|
|
|
|
|
@param[in] thd Thread context.
|
|
|
|
@param[in] prelocking_ctx Prelocking context of the statement.
|
|
|
|
@param[in] table_list Table list element for table.
|
|
|
|
@param[in] sp Routine body.
|
|
|
|
@param[out] need_prelocking Set to TRUE if method detects that prelocking
|
|
|
|
required, not changed otherwise.
|
|
|
|
|
|
|
|
|
|
|
|
@retval FALSE Success.
|
|
|
|
@retval TRUE Failure (OOM).
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool Alter_table_prelocking_strategy::
|
|
|
|
handle_table(THD *thd, Query_tables_list *prelocking_ctx,
|
|
|
|
TABLE_LIST *table_list, bool *need_prelocking)
|
|
|
|
{
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Defines how prelocking algorithm for ALTER TABLE statement
|
|
|
|
should handle view - do nothing. We don't need to add view
|
|
|
|
routines to the prelocking set in this case as view is not going
|
|
|
|
to be materialized.
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool Alter_table_prelocking_strategy::
|
|
|
|
handle_view(THD *thd, Query_tables_list *prelocking_ctx,
|
|
|
|
TABLE_LIST *table_list, bool *need_prelocking)
|
|
|
|
{
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2002-11-16 19:19:10 +01:00
|
|
|
/*
|
|
|
|
Check that lock is ok for tables; Call start stmt if ok
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
check_lock_and_start_stmt()
|
|
|
|
thd Thread handle
|
|
|
|
table_list Table to check
|
|
|
|
lock_type Lock used for table
|
|
|
|
|
|
|
|
RETURN VALUES
|
|
|
|
0 ok
|
|
|
|
1 error
|
|
|
|
*/
|
|
|
|
|
|
|
|
static bool check_lock_and_start_stmt(THD *thd, TABLE *table,
|
|
|
|
thr_lock_type lock_type)
|
|
|
|
{
|
|
|
|
int error;
|
|
|
|
DBUG_ENTER("check_lock_and_start_stmt");
|
|
|
|
|
|
|
|
if ((int) lock_type >= (int) TL_WRITE_ALLOW_READ &&
|
|
|
|
(int) table->reginfo.lock_type < (int) TL_WRITE_ALLOW_READ)
|
|
|
|
{
|
2005-01-06 12:00:13 +01:00
|
|
|
my_error(ER_TABLE_NOT_LOCKED_FOR_WRITE, MYF(0),table->alias);
|
2002-11-16 19:19:10 +01:00
|
|
|
DBUG_RETURN(1);
|
|
|
|
}
|
2005-09-30 20:20:10 +02:00
|
|
|
if ((error=table->file->start_stmt(thd, lock_type)))
|
2002-11-16 19:19:10 +01:00
|
|
|
{
|
|
|
|
table->file->print_error(error,MYF(0));
|
|
|
|
DBUG_RETURN(1);
|
|
|
|
}
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
/**
|
|
|
|
@brief Open and lock one table
|
|
|
|
|
|
|
|
@param[in] thd thread handle
|
|
|
|
@param[in] table_l table to open is first table in this list
|
|
|
|
@param[in] lock_type lock to use for table
|
2009-12-01 14:38:00 +01:00
|
|
|
@param[in] flags options to be used while opening and locking
|
|
|
|
table (see open_table(), mysql_lock_tables())
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
|
|
|
|
@return table
|
|
|
|
@retval != NULL OK, opened table returned
|
|
|
|
@retval NULL Error
|
|
|
|
|
|
|
|
@note
|
|
|
|
If ok, the following are also set:
|
|
|
|
table_list->lock_type lock_type
|
|
|
|
table_list->table table
|
|
|
|
|
|
|
|
@note
|
|
|
|
If table_l is a list, not a single table, the list is temporarily
|
|
|
|
broken.
|
|
|
|
|
|
|
|
@detail
|
|
|
|
This function is meant as a replacement for open_ltable() when
|
|
|
|
MERGE tables can be opened. open_ltable() cannot open MERGE tables.
|
|
|
|
|
|
|
|
There may be more differences between open_n_lock_single_table() and
|
|
|
|
open_ltable(). One known difference is that open_ltable() does
|
|
|
|
neither call decide_logging_format() nor handle some other logging
|
|
|
|
and locking issues because it does not call lock_tables().
|
|
|
|
*/
|
|
|
|
|
|
|
|
TABLE *open_n_lock_single_table(THD *thd, TABLE_LIST *table_l,
|
2009-12-01 14:38:00 +01:00
|
|
|
thr_lock_type lock_type, uint flags)
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
{
|
|
|
|
TABLE_LIST *save_next_global;
|
|
|
|
DBUG_ENTER("open_n_lock_single_table");
|
|
|
|
|
|
|
|
/* Remember old 'next' pointer. */
|
|
|
|
save_next_global= table_l->next_global;
|
|
|
|
/* Break list. */
|
|
|
|
table_l->next_global= NULL;
|
|
|
|
|
|
|
|
/* Set requested lock type. */
|
|
|
|
table_l->lock_type= lock_type;
|
|
|
|
/* Allow to open real tables only. */
|
|
|
|
table_l->required_type= FRMTYPE_TABLE;
|
|
|
|
|
|
|
|
/* Open the table. */
|
2009-12-01 14:38:00 +01:00
|
|
|
if (open_and_lock_tables_derived(thd, table_l, FALSE, flags))
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
table_l->table= NULL; /* Just to be sure. */
|
|
|
|
|
|
|
|
/* Restore list. */
|
|
|
|
table_l->next_global= save_next_global;
|
|
|
|
|
|
|
|
DBUG_RETURN(table_l->table);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2002-11-16 19:19:10 +01:00
|
|
|
/*
|
|
|
|
Open and lock one table
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
open_ltable()
|
|
|
|
thd Thread handler
|
|
|
|
table_list Table to open is first table in this list
|
|
|
|
lock_type Lock to use for open
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
lock_flags Flags passed to mysql_lock_table
|
2002-11-16 19:19:10 +01:00
|
|
|
|
2005-03-04 14:35:28 +01:00
|
|
|
NOTE
|
2009-03-27 12:09:15 +01:00
|
|
|
This function doesn't do anything like SP/SF/views/triggers analysis done
|
2009-03-27 13:55:14 +01:00
|
|
|
in open_table()/lock_tables(). It is intended for opening of only one
|
2009-03-27 12:09:15 +01:00
|
|
|
concrete table. And used only in special contexts.
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2002-11-16 19:19:10 +01:00
|
|
|
RETURN VALUES
|
|
|
|
table Opened table
|
|
|
|
0 Error
|
|
|
|
|
|
|
|
If ok, the following are also set:
|
|
|
|
table_list->lock_type lock_type
|
|
|
|
table_list->table table
|
|
|
|
*/
|
|
|
|
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
TABLE *open_ltable(THD *thd, TABLE_LIST *table_list, thr_lock_type lock_type,
|
|
|
|
uint lock_flags)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
TABLE *table;
|
2009-12-08 10:57:07 +01:00
|
|
|
Open_table_context ot_ctx(thd);
|
2000-07-31 21:29:14 +02:00
|
|
|
bool refresh;
|
2009-12-01 15:58:31 +01:00
|
|
|
bool error;
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("open_ltable");
|
|
|
|
|
2009-03-27 13:55:14 +01:00
|
|
|
/* should not be used in a prelocked_mode context, see NOTE above */
|
2009-12-01 15:39:03 +01:00
|
|
|
DBUG_ASSERT(thd->locked_tables_mode < LTM_PRELOCKED);
|
2009-03-27 13:55:14 +01:00
|
|
|
|
2007-02-22 16:03:08 +01:00
|
|
|
thd_proc_info(thd, "Opening table");
|
2003-08-19 15:00:12 +02:00
|
|
|
thd->current_tablenr= 0;
|
2004-08-24 14:37:51 +02:00
|
|
|
/* open_ltable can be used only for BASIC TABLEs */
|
|
|
|
table_list->required_type= FRMTYPE_TABLE;
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
retry:
|
2009-12-08 10:57:07 +01:00
|
|
|
while ((error= open_table(thd, table_list, thd->mem_root, &ot_ctx, 0)) &&
|
|
|
|
ot_ctx.can_recover_from_failed_open_table())
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
|
|
|
/*
|
2009-12-08 10:57:07 +01:00
|
|
|
Even though we have failed to open table we still need to
|
|
|
|
call release_all_locks() to release metadata locks which
|
2009-11-30 16:55:03 +01:00
|
|
|
might have been acquired successfully.
|
|
|
|
*/
|
2009-12-08 10:57:07 +01:00
|
|
|
if (! thd->locked_tables_mode)
|
|
|
|
thd->mdl_context.release_all_locks();
|
|
|
|
table_list->mdl_request.ticket= 0;
|
|
|
|
if (ot_ctx.recover_from_failed_open_table_attempt(thd, table_list))
|
2009-11-30 16:55:03 +01:00
|
|
|
break;
|
|
|
|
}
|
2002-11-23 22:21:49 +01:00
|
|
|
|
2009-12-01 15:58:31 +01:00
|
|
|
if (!error)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2009-12-01 15:58:31 +01:00
|
|
|
/*
|
Backport of revno ## 2617.31.1, 2617.31.3, 2617.31.4, 2617.31.5,
2617.31.12, 2617.31.15, 2617.31.15, 2617.31.16, 2617.43.1
- initial changeset that introduced the fix for
Bug#989 and follow up fixes for all test suite failures
introduced in the initial changeset.
------------------------------------------------------------
revno: 2617.31.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Fri 2009-03-06 19:17:00 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
Currently the MySQL server does not keep metadata locks on
schema objects for the duration of a transaction, thus failing
to guarantee the integrity of the schema objects being used
during the transaction and to protect then from concurrent
DDL operations. This also poses a problem for replication as
a DDL operation might be replicated even thought there are
active transactions using the object being modified.
The solution is to defer the release of metadata locks until
a active transaction is either committed or rolled back. This
prevents other statements from modifying the table for the
entire duration of the transaction. This provides commitment
ordering for guaranteeing serializability across multiple
transactions.
- Incompatible change:
If MySQL's metadata locking system encounters a lock conflict,
the usual schema is to use the try and back-off technique to
avoid deadlocks -- this schema consists in releasing all locks
and trying to acquire them all in one go.
But in a transactional context this algorithm can't be utilized
as its not possible to release locks acquired during the course
of the transaction without breaking the transaction commitments.
To avoid deadlocks in this case, the ER_LOCK_DEADLOCK will be
returned if a lock conflict is encountered during a transaction.
Let's consider an example:
A transaction has two statements that modify table t1, then table
t2, and then commits. The first statement of the transaction will
acquire a shared metadata lock on table t1, and it will be kept
utill COMMIT to ensure serializability.
At the moment when the second statement attempts to acquire a
shared metadata lock on t2, a concurrent ALTER or DROP statement
might have locked t2 exclusively. The prescription of the current
locking protocol is that the acquirer of the shared lock backs off
-- gives up all his current locks and retries. This implies that
the entire multi-statement transaction has to be rolled back.
- Incompatible change:
FLUSH commands such as FLUSH PRIVILEGES and FLUSH TABLES WITH READ
LOCK won't cause locked tables to be implicitly unlocked anymore.
2009-12-05 00:02:48 +01:00
|
|
|
We can't have a view or some special "open_strategy" in this function
|
2009-12-01 15:58:31 +01:00
|
|
|
so there should be a TABLE instance.
|
|
|
|
*/
|
|
|
|
DBUG_ASSERT(table_list->table);
|
|
|
|
table= table_list->table;
|
2009-12-03 00:09:22 +01:00
|
|
|
if (table->file->ht->db_type == DB_TYPE_MRG_MYISAM)
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
{
|
|
|
|
/* A MERGE table must not come here. */
|
|
|
|
/* purecov: begin tested */
|
|
|
|
my_error(ER_WRONG_OBJECT, MYF(0), table->s->db.str,
|
|
|
|
table->s->table_name.str, "BASE TABLE");
|
|
|
|
table= 0;
|
|
|
|
goto end;
|
|
|
|
/* purecov: end */
|
|
|
|
}
|
|
|
|
|
2002-11-16 19:19:10 +01:00
|
|
|
table_list->lock_type= lock_type;
|
2000-07-31 21:29:14 +02:00
|
|
|
table->grant= table_list->grant;
|
2009-12-01 15:39:03 +01:00
|
|
|
if (thd->locked_tables_mode)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2002-11-16 19:19:10 +01:00
|
|
|
if (check_lock_and_start_stmt(thd, table, lock_type))
|
|
|
|
table= 0;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2003-02-12 20:55:37 +01:00
|
|
|
DBUG_ASSERT(thd->lock == 0); // You must lock everything at once
|
2002-11-16 19:19:10 +01:00
|
|
|
if ((table->reginfo.lock_type= lock_type) != TL_UNLOCK)
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
if (! (thd->lock= mysql_lock_tables(thd, &table_list->table, 1,
|
2009-11-30 20:03:37 +01:00
|
|
|
lock_flags, &refresh)))
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
|
|
|
if (refresh)
|
|
|
|
{
|
2009-12-08 10:57:07 +01:00
|
|
|
if (ot_ctx.can_deadlock())
|
|
|
|
{
|
|
|
|
my_error(ER_LOCK_DEADLOCK, MYF(0));
|
|
|
|
table= 0;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
close_thread_tables(thd);
|
|
|
|
table_list->table= NULL;
|
|
|
|
table_list->mdl_request.ticket= NULL;
|
|
|
|
if (! thd->locked_tables_mode)
|
|
|
|
thd->mdl_context.release_all_locks();
|
|
|
|
goto retry;
|
|
|
|
}
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
|
|
|
else
|
|
|
|
table= 0;
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
}
|
2009-12-01 15:58:31 +01:00
|
|
|
else
|
|
|
|
table= 0;
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
end:
|
2007-02-22 16:03:08 +01:00
|
|
|
thd_proc_info(thd, 0);
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_RETURN(table);
|
|
|
|
}
|
|
|
|
|
2002-11-16 19:19:10 +01:00
|
|
|
|
2009-12-08 15:13:12 +01:00
|
|
|
/**
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
Open all tables in list, locks them and optionally process derived tables.
|
2004-06-25 15:52:01 +02:00
|
|
|
|
2009-12-08 15:13:12 +01:00
|
|
|
@param thd Thread context.
|
|
|
|
@param tables List of tables for open and locking.
|
|
|
|
@param derived If to handle derived tables.
|
|
|
|
@param flags Bitmap of options to be used to open and lock
|
|
|
|
tables (see open_tables() and mysql_lock_tables()
|
|
|
|
for details).
|
|
|
|
@param prelocking_strategy Strategy which specifies how prelocking algorithm
|
|
|
|
should work for this statement.
|
2004-06-25 15:52:01 +02:00
|
|
|
|
2009-12-08 15:13:12 +01:00
|
|
|
@note
|
2005-01-16 13:16:23 +01:00
|
|
|
The lock will automaticaly be freed by close_thread_tables()
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
|
2009-12-08 15:13:12 +01:00
|
|
|
@note
|
|
|
|
There are several convenience functions, e.g. :
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
- simple_open_n_lock_tables(thd, tables) without derived handling
|
|
|
|
- open_and_lock_tables(thd, tables) with derived handling
|
|
|
|
Both inline functions call open_and_lock_tables_derived() with
|
|
|
|
the third argument set appropriately.
|
2009-12-08 15:13:12 +01:00
|
|
|
|
|
|
|
@retval FALSE OK.
|
|
|
|
@retval TRUE Error
|
2004-06-25 15:52:01 +02:00
|
|
|
*/
|
|
|
|
|
2009-12-09 10:04:55 +01:00
|
|
|
bool open_and_lock_tables_derived(THD *thd, TABLE_LIST *tables,
|
|
|
|
bool derived, uint flags,
|
|
|
|
Prelocking_strategy *prelocking_strategy)
|
2004-06-25 15:52:01 +02:00
|
|
|
{
|
|
|
|
uint counter;
|
2005-09-15 01:56:09 +02:00
|
|
|
bool need_reopen;
|
2009-12-08 10:57:07 +01:00
|
|
|
bool has_locks= thd->mdl_context.has_locks();
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
DBUG_ENTER("open_and_lock_tables_derived");
|
|
|
|
DBUG_PRINT("enter", ("derived handling: %d", derived));
|
2005-09-15 01:56:09 +02:00
|
|
|
|
|
|
|
for ( ; ; )
|
|
|
|
{
|
2009-12-08 15:13:12 +01:00
|
|
|
if (open_tables(thd, &tables, &counter, flags, prelocking_strategy))
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
DBUG_EXECUTE_IF("sleep_open_and_lock_after_open", {
|
|
|
|
const char *old_proc_info= thd->proc_info;
|
|
|
|
thd->proc_info= "DBUG sleep";
|
|
|
|
my_sleep(6000000);
|
|
|
|
thd->proc_info= old_proc_info;});
|
|
|
|
|
2009-11-30 20:03:37 +01:00
|
|
|
if (!lock_tables(thd, tables, counter, flags,
|
|
|
|
&need_reopen))
|
2005-09-15 01:56:09 +02:00
|
|
|
break;
|
|
|
|
if (!need_reopen)
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2009-12-08 10:57:07 +01:00
|
|
|
if (thd->in_multi_stmt_transaction() && has_locks)
|
|
|
|
{
|
|
|
|
my_error(ER_LOCK_DEADLOCK, MYF(0));
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2009-12-08 10:57:07 +01:00
|
|
|
}
|
|
|
|
close_tables_for_reopen(thd, &tables);
|
2005-09-15 01:56:09 +02:00
|
|
|
}
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
if (derived &&
|
|
|
|
(mysql_handle_derived(thd->lex, &mysql_derived_prepare) ||
|
|
|
|
(thd->fill_derived_tables() &&
|
|
|
|
mysql_handle_derived(thd->lex, &mysql_derived_filling))))
|
2004-10-20 03:04:37 +02:00
|
|
|
DBUG_RETURN(TRUE); /* purecov: inspected */
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_RETURN(FALSE);
|
2004-10-07 09:50:13 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-01-13 23:09:15 +01:00
|
|
|
/*
|
|
|
|
Open all tables in list and process derived tables
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
open_normal_and_derived_tables
|
|
|
|
thd - thread handler
|
2005-06-03 00:15:56 +02:00
|
|
|
tables - list of tables for open
|
2005-08-08 15:46:06 +02:00
|
|
|
flags - bitmap of flags to modify how the tables will be open:
|
|
|
|
MYSQL_LOCK_IGNORE_FLUSH - open table even if someone has
|
|
|
|
done a flush or namelock on it.
|
2005-01-13 23:09:15 +01:00
|
|
|
|
|
|
|
RETURN
|
|
|
|
FALSE - ok
|
|
|
|
TRUE - error
|
|
|
|
|
|
|
|
NOTE
|
|
|
|
This is to be used on prepare stage when you don't read any
|
|
|
|
data from the tables.
|
|
|
|
*/
|
|
|
|
|
2005-08-08 15:46:06 +02:00
|
|
|
bool open_normal_and_derived_tables(THD *thd, TABLE_LIST *tables, uint flags)
|
2005-01-13 23:09:15 +01:00
|
|
|
{
|
|
|
|
uint counter;
|
|
|
|
DBUG_ENTER("open_normal_and_derived_tables");
|
2005-01-15 16:38:43 +01:00
|
|
|
DBUG_ASSERT(!thd->fill_derived_tables());
|
2005-08-08 15:46:06 +02:00
|
|
|
if (open_tables(thd, &tables, &counter, flags) ||
|
2005-01-15 16:38:43 +01:00
|
|
|
mysql_handle_derived(thd->lex, &mysql_derived_prepare))
|
|
|
|
DBUG_RETURN(TRUE); /* purecov: inspected */
|
|
|
|
DBUG_RETURN(0);
|
2005-01-13 23:09:15 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-03-04 14:35:28 +01:00
|
|
|
/*
|
|
|
|
Mark all real tables in the list as free for reuse.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
mark_real_tables_as_free_for_reuse()
|
|
|
|
thd - thread context
|
|
|
|
table - head of the list of tables
|
|
|
|
|
|
|
|
DESCRIPTION
|
|
|
|
Marks all real tables in the list (i.e. not views, derived
|
|
|
|
or schema tables) as free for reuse.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static void mark_real_tables_as_free_for_reuse(TABLE_LIST *table)
|
|
|
|
{
|
|
|
|
for (; table; table= table->next_global)
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
if (!table->placeholder())
|
2005-03-04 14:35:28 +01:00
|
|
|
table->table->query_id= 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-05-29 17:13:17 +02:00
|
|
|
/**
|
|
|
|
Decide on logging format to use for the statement.
|
2002-11-16 19:19:10 +01:00
|
|
|
|
2007-05-29 17:13:17 +02:00
|
|
|
Compute the capabilities vector for the involved storage engines
|
|
|
|
and mask out the flags for the binary log. Right now, the binlog
|
|
|
|
flags only include the capabilities of the storage engines, so this
|
|
|
|
is safe.
|
2002-11-16 19:19:10 +01:00
|
|
|
|
2007-05-29 17:13:17 +02:00
|
|
|
We now have three alternatives that prevent the statement from
|
|
|
|
being loggable:
|
2003-02-12 20:55:37 +01:00
|
|
|
|
2007-05-29 17:13:17 +02:00
|
|
|
1. If there are no capabilities left (all flags are clear) it is
|
|
|
|
not possible to log the statement at all, so we roll back the
|
|
|
|
statement and report an error.
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2007-05-29 17:13:17 +02:00
|
|
|
2. Statement mode is set, but the capabilities indicate that
|
|
|
|
statement format is not possible.
|
2002-11-16 19:19:10 +01:00
|
|
|
|
2007-05-29 17:13:17 +02:00
|
|
|
3. Row mode is set, but the capabilities indicate that row
|
|
|
|
format is not possible.
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2007-05-29 17:13:17 +02:00
|
|
|
4. Statement is unsafe, but the capabilities indicate that row
|
|
|
|
format is not possible.
|
|
|
|
|
|
|
|
If we are in MIXED mode, we then decide what logging format to use:
|
|
|
|
|
|
|
|
1. If the statement is unsafe, row-based logging is used.
|
2005-09-15 01:56:09 +02:00
|
|
|
|
2007-05-29 17:13:17 +02:00
|
|
|
2. If statement-based logging is not possible, row-based logging is
|
|
|
|
used.
|
|
|
|
|
|
|
|
3. Otherwise, statement-based logging is used.
|
|
|
|
|
|
|
|
@param thd Client thread
|
|
|
|
@param tables Tables involved in the query
|
|
|
|
*/
|
|
|
|
|
2009-12-09 10:04:55 +01:00
|
|
|
bool decide_logging_format(THD *thd, TABLE_LIST *tables)
|
2007-05-29 17:13:17 +02:00
|
|
|
{
|
2009-09-24 16:52:52 +02:00
|
|
|
/*
|
|
|
|
In SBR mode, we are only proceeding if we are binlogging this
|
|
|
|
statement, ie, the filtering rules won't later filter this out.
|
|
|
|
|
|
|
|
This check here is needed to prevent some spurious error to be
|
|
|
|
raised in some cases (See BUG#42829).
|
|
|
|
*/
|
|
|
|
if (mysql_bin_log.is_open() && (thd->options & OPTION_BIN_LOG) &&
|
|
|
|
(thd->variables.binlog_format != BINLOG_FORMAT_STMT ||
|
|
|
|
binlog_filter->db_ok(thd->db)))
|
2007-05-28 12:50:29 +02:00
|
|
|
{
|
2009-02-06 17:06:41 +01:00
|
|
|
/*
|
|
|
|
Compute the starting vectors for the computations by creating a
|
|
|
|
set with all the capabilities bits set and one with no
|
|
|
|
capabilities bits set.
|
|
|
|
*/
|
|
|
|
handler::Table_flags flags_some_set= 0;
|
|
|
|
handler::Table_flags flags_all_set=
|
|
|
|
HA_BINLOG_ROW_CAPABLE | HA_BINLOG_STMT_CAPABLE;
|
|
|
|
|
2007-06-21 13:34:06 +02:00
|
|
|
my_bool multi_engine= FALSE;
|
|
|
|
void* prev_ht= NULL;
|
2007-05-29 17:13:17 +02:00
|
|
|
for (TABLE_LIST *table= tables; table; table= table->next_global)
|
2007-06-21 13:34:06 +02:00
|
|
|
{
|
2008-09-29 15:53:40 +02:00
|
|
|
if (table->placeholder())
|
|
|
|
continue;
|
|
|
|
if (table->table->s->table_category == TABLE_CATEGORY_PERFORMANCE)
|
|
|
|
thd->lex->set_stmt_unsafe();
|
|
|
|
if (table->lock_type >= TL_WRITE_ALLOW_WRITE)
|
2007-05-28 12:50:29 +02:00
|
|
|
{
|
2007-06-21 13:34:06 +02:00
|
|
|
ulonglong const flags= table->table->file->ha_table_flags();
|
2007-05-29 17:13:17 +02:00
|
|
|
DBUG_PRINT("info", ("table: %s; ha_table_flags: %s%s",
|
2007-05-28 12:50:29 +02:00
|
|
|
table->table_name,
|
2007-05-29 17:13:17 +02:00
|
|
|
FLAGSTR(flags, HA_BINLOG_STMT_CAPABLE),
|
|
|
|
FLAGSTR(flags, HA_BINLOG_ROW_CAPABLE)));
|
2007-06-21 13:34:06 +02:00
|
|
|
if (prev_ht && prev_ht != table->table->file->ht)
|
|
|
|
multi_engine= TRUE;
|
|
|
|
prev_ht= table->table->file->ht;
|
|
|
|
flags_all_set &= flags;
|
|
|
|
flags_some_set |= flags;
|
2007-05-28 12:50:29 +02:00
|
|
|
}
|
2007-06-21 13:34:06 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
DBUG_PRINT("info", ("flags_all_set: %s%s",
|
|
|
|
FLAGSTR(flags_all_set, HA_BINLOG_STMT_CAPABLE),
|
|
|
|
FLAGSTR(flags_all_set, HA_BINLOG_ROW_CAPABLE)));
|
|
|
|
DBUG_PRINT("info", ("flags_some_set: %s%s",
|
|
|
|
FLAGSTR(flags_some_set, HA_BINLOG_STMT_CAPABLE),
|
|
|
|
FLAGSTR(flags_some_set, HA_BINLOG_ROW_CAPABLE)));
|
2007-05-28 12:50:29 +02:00
|
|
|
DBUG_PRINT("info", ("thd->variables.binlog_format: %ld",
|
|
|
|
thd->variables.binlog_format));
|
2007-06-21 13:34:06 +02:00
|
|
|
DBUG_PRINT("info", ("multi_engine: %s",
|
|
|
|
multi_engine ? "TRUE" : "FALSE"));
|
2007-05-28 12:50:29 +02:00
|
|
|
|
|
|
|
int error= 0;
|
2007-06-21 13:34:06 +02:00
|
|
|
if (flags_all_set == 0)
|
2007-05-28 12:50:29 +02:00
|
|
|
{
|
2007-06-12 21:47:00 +02:00
|
|
|
my_error((error= ER_BINLOG_LOGGING_IMPOSSIBLE), MYF(0),
|
|
|
|
"Statement cannot be logged to the binary log in"
|
|
|
|
" row-based nor statement-based format");
|
2007-05-28 12:50:29 +02:00
|
|
|
}
|
|
|
|
else if (thd->variables.binlog_format == BINLOG_FORMAT_STMT &&
|
2007-06-21 13:34:06 +02:00
|
|
|
(flags_all_set & HA_BINLOG_STMT_CAPABLE) == 0)
|
2007-05-28 12:50:29 +02:00
|
|
|
{
|
2007-06-12 21:47:00 +02:00
|
|
|
my_error((error= ER_BINLOG_LOGGING_IMPOSSIBLE), MYF(0),
|
|
|
|
"Statement-based format required for this statement,"
|
|
|
|
" but not allowed by this combination of engines");
|
2007-05-28 12:50:29 +02:00
|
|
|
}
|
|
|
|
else if ((thd->variables.binlog_format == BINLOG_FORMAT_ROW ||
|
|
|
|
thd->lex->is_stmt_unsafe()) &&
|
2007-06-21 13:34:06 +02:00
|
|
|
(flags_all_set & HA_BINLOG_ROW_CAPABLE) == 0)
|
2007-05-28 12:50:29 +02:00
|
|
|
{
|
2007-06-12 21:47:00 +02:00
|
|
|
my_error((error= ER_BINLOG_LOGGING_IMPOSSIBLE), MYF(0),
|
|
|
|
"Row-based format required for this statement,"
|
|
|
|
" but not allowed by this combination of engines");
|
2007-05-28 12:50:29 +02:00
|
|
|
}
|
|
|
|
|
2007-06-21 13:34:06 +02:00
|
|
|
/*
|
|
|
|
If more than one engine is involved in the statement and at
|
|
|
|
least one is doing it's own logging (is *self-logging*), the
|
|
|
|
statement cannot be logged atomically, so we generate an error
|
|
|
|
rather than allowing the binlog to become corrupt.
|
|
|
|
*/
|
|
|
|
if (multi_engine &&
|
|
|
|
(flags_some_set & HA_HAS_OWN_BINLOGGING))
|
|
|
|
{
|
|
|
|
error= ER_BINLOG_LOGGING_IMPOSSIBLE;
|
|
|
|
my_error(error, MYF(0),
|
|
|
|
"Statement cannot be written atomically since more"
|
|
|
|
" than one engine involved and at least one engine"
|
|
|
|
" is self-logging");
|
|
|
|
}
|
|
|
|
|
2007-05-28 12:50:29 +02:00
|
|
|
DBUG_PRINT("info", ("error: %d", error));
|
|
|
|
|
|
|
|
if (error)
|
2009-12-09 10:04:55 +01:00
|
|
|
return TRUE;
|
2007-05-28 12:50:29 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
We switch to row-based format if we are in mixed mode and one of
|
|
|
|
the following are true:
|
|
|
|
|
|
|
|
1. If the statement is unsafe
|
|
|
|
2. If statement format cannot be used
|
|
|
|
|
|
|
|
Observe that point to cannot be decided before the tables
|
|
|
|
involved in a statement has been checked, i.e., we cannot put
|
|
|
|
this code in reset_current_stmt_binlog_row_based(), it has to be
|
|
|
|
here.
|
2007-05-29 17:13:17 +02:00
|
|
|
*/
|
2007-05-28 12:50:29 +02:00
|
|
|
if (thd->lex->is_stmt_unsafe() ||
|
2007-06-21 13:34:06 +02:00
|
|
|
(flags_all_set & HA_BINLOG_STMT_CAPABLE) == 0)
|
2007-05-28 12:50:29 +02:00
|
|
|
{
|
|
|
|
thd->set_current_stmt_binlog_row_based_if_mixed();
|
|
|
|
}
|
|
|
|
}
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
|
2009-12-09 10:04:55 +01:00
|
|
|
return FALSE;
|
2007-05-29 17:13:17 +02:00
|
|
|
}
|
|
|
|
|
2002-11-16 19:19:10 +01:00
|
|
|
/*
|
|
|
|
Lock all tables in list
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
lock_tables()
|
|
|
|
thd Thread handler
|
|
|
|
tables Tables to lock
|
2005-09-15 01:56:09 +02:00
|
|
|
count Number of opened tables
|
2009-11-30 20:03:37 +01:00
|
|
|
flags Options (see mysql_lock_tables() for details)
|
2005-09-15 01:56:09 +02:00
|
|
|
need_reopen Out parameter which if TRUE indicates that some
|
|
|
|
tables were dropped or altered during this call
|
|
|
|
and therefore invoker should reopen tables and
|
|
|
|
try to lock them once again (in this case
|
|
|
|
lock_tables() will also return error).
|
2002-11-16 19:19:10 +01:00
|
|
|
|
2003-02-12 20:55:37 +01:00
|
|
|
NOTES
|
|
|
|
You can't call lock_tables twice, as this would break the dead-lock-free
|
|
|
|
handling thr_lock gives us. You most always get all needed locks at
|
|
|
|
once.
|
|
|
|
|
2009-12-01 15:39:03 +01:00
|
|
|
If query for which we are calling this function marked as requiring
|
|
|
|
prelocking, this function will change locked_tables_mode to
|
|
|
|
LTM_PRELOCKED.
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2002-11-16 19:19:10 +01:00
|
|
|
RETURN VALUES
|
|
|
|
0 ok
|
|
|
|
-1 Error
|
|
|
|
*/
|
|
|
|
|
2009-12-09 10:04:55 +01:00
|
|
|
bool lock_tables(THD *thd, TABLE_LIST *tables, uint count,
|
|
|
|
uint flags, bool *need_reopen)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2001-04-19 19:41:19 +02:00
|
|
|
TABLE_LIST *table;
|
2005-03-04 14:35:28 +01:00
|
|
|
|
|
|
|
DBUG_ENTER("lock_tables");
|
|
|
|
/*
|
|
|
|
We can't meet statement requiring prelocking if we already
|
|
|
|
in prelocked mode.
|
|
|
|
*/
|
2009-12-01 15:39:03 +01:00
|
|
|
DBUG_ASSERT(thd->locked_tables_mode <= LTM_LOCK_TABLES ||
|
|
|
|
!thd->lex->requires_prelocking());
|
2005-09-15 01:56:09 +02:00
|
|
|
*need_reopen= FALSE;
|
|
|
|
|
2006-07-26 12:40:26 +02:00
|
|
|
if (!tables && !thd->lex->requires_prelocking())
|
2007-05-29 17:13:17 +02:00
|
|
|
DBUG_RETURN(decide_logging_format(thd, tables));
|
2002-11-16 19:19:10 +01:00
|
|
|
|
2005-03-04 14:35:28 +01:00
|
|
|
/*
|
2009-12-01 15:39:03 +01:00
|
|
|
Check for thd->locked_tables_mode to avoid a redundant
|
|
|
|
and harmful attempt to lock the already locked tables again.
|
|
|
|
Checking for thd->lock is not enough in some situations. For example,
|
|
|
|
if a stored function contains
|
2005-03-04 14:35:28 +01:00
|
|
|
"drop table t3; create temporary t3 ..; insert into t3 ...;"
|
2009-12-01 15:39:03 +01:00
|
|
|
thd->lock may be 0 after drop tables, whereas locked_tables_mode
|
|
|
|
is still on. In this situation an attempt to lock temporary
|
|
|
|
table t3 will lead to a memory leak.
|
2005-03-04 14:35:28 +01:00
|
|
|
*/
|
2009-12-01 15:39:03 +01:00
|
|
|
if (! thd->locked_tables_mode)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2003-02-12 20:55:37 +01:00
|
|
|
DBUG_ASSERT(thd->lock == 0); // You must lock everything at once
|
2000-07-31 21:29:14 +02:00
|
|
|
TABLE **start,**ptr;
|
2005-03-04 14:35:28 +01:00
|
|
|
|
2005-01-06 12:00:13 +01:00
|
|
|
if (!(ptr=start=(TABLE**) thd->alloc(sizeof(TABLE*)*count)))
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2004-07-16 00:15:55 +02:00
|
|
|
for (table= tables; table; table= table->next_global)
|
2004-06-25 15:52:01 +02:00
|
|
|
{
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
if (!table->placeholder())
|
2004-06-25 15:52:01 +02:00
|
|
|
*(ptr++)= table->table;
|
|
|
|
}
|
2005-03-04 14:35:28 +01:00
|
|
|
|
|
|
|
/* We have to emulate LOCK TABLES if we are statement needs prelocking. */
|
|
|
|
if (thd->lex->requires_prelocking())
|
|
|
|
{
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
/*
|
2009-10-01 01:19:36 +02:00
|
|
|
A query that modifies autoinc column in sub-statement can make the
|
|
|
|
master and slave inconsistent.
|
|
|
|
We can solve these problems in mixed mode by switching to binlogging
|
|
|
|
if at least one updated table is used by sub-statement
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
*/
|
2009-11-13 09:29:35 +01:00
|
|
|
if (thd->variables.binlog_format != BINLOG_FORMAT_ROW && tables &&
|
2009-10-01 01:19:36 +02:00
|
|
|
has_write_table_with_auto_increment(thd->lex->first_not_own_table()))
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
{
|
2007-05-14 14:45:38 +02:00
|
|
|
thd->lex->set_stmt_unsafe();
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
}
|
2005-03-04 14:35:28 +01:00
|
|
|
}
|
|
|
|
|
2009-09-29 17:38:40 +02:00
|
|
|
DEBUG_SYNC(thd, "before_lock_tables_takes_lock");
|
|
|
|
|
2005-09-15 01:56:09 +02:00
|
|
|
if (! (thd->lock= mysql_lock_tables(thd, start, (uint) (ptr - start),
|
2009-11-30 20:03:37 +01:00
|
|
|
flags, need_reopen)))
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2007-05-29 17:13:17 +02:00
|
|
|
|
2005-03-04 14:35:28 +01:00
|
|
|
if (thd->lex->requires_prelocking() &&
|
|
|
|
thd->lex->sql_command != SQLCOM_LOCK_TABLES)
|
|
|
|
{
|
|
|
|
TABLE_LIST *first_not_own= thd->lex->first_not_own_table();
|
|
|
|
/*
|
|
|
|
We just have done implicit LOCK TABLES, and now we have
|
|
|
|
to emulate first open_and_lock_tables() after it.
|
|
|
|
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
When open_and_lock_tables() is called for a single table out of
|
|
|
|
a table list, the 'next_global' chain is temporarily broken. We
|
|
|
|
may not find 'first_not_own' before the end of the "list".
|
|
|
|
Look for example at those places where open_n_lock_single_table()
|
|
|
|
is called. That function implements the temporary breaking of
|
|
|
|
a table list for opening a single table.
|
|
|
|
*/
|
|
|
|
for (table= tables;
|
|
|
|
table && table != first_not_own;
|
|
|
|
table= table->next_global)
|
2005-03-04 14:35:28 +01:00
|
|
|
{
|
Bug#8407 (Stored functions/triggers ignore exception handler)
Bug 18914 (Calling certain SPs from triggers fail)
Bug 20713 (Functions will not not continue for SQLSTATE VALUE '42S02')
Bug 21825 (Incorrect message error deleting records in a table with a
trigger for inserting)
Bug 22580 (DROP TABLE in nested stored procedure causes strange dependency
error)
Bug 25345 (Cursors from Functions)
This fix resolves a long standing issue originally reported with bug 8407,
which affect the behavior of Stored Procedures, Stored Functions and Trigger
in many different ways, causing symptoms reported by all the bugs listed.
In all cases, the root cause of the problem traces back to 8407 and how the
server locks tables involved with sub statements.
Prior to this fix, the implementation of stored routines would:
- compute the transitive closure of all the tables referenced by a top level
statement
- open and lock all the tables involved
- execute the top level statement
"transitive closure of tables" means collecting:
- all the tables,
- all the stored functions,
- all the views,
- all the table triggers
- all the stored procedures
involved, and recursively inspect these objects definition to find more
references to more objects, until the list of every object referenced does
not grow any more.
This mechanism is known as "pre-locking" tables before execution.
The motivation for locking all the tables (possibly) used at once is to
prevent dead locks.
One problem with this approach is that, if the execution path the code
really takes during runtime does not use a given table, and if the table is
missing, the server would not execute the statement.
This in particular has a major impact on triggers, since a missing table
referenced by an update/delete trigger would prevent an insert trigger to run.
Another problem is that stored routines might define SQL exception handlers
to deal with missing tables, but the server implementation would never give
user code a chance to execute this logic, since the routine is never
executed when a missing table cause the pre-locking code to fail.
With this fix, the internal implementation of the pre-locking code has been
relaxed of some constraints, so that failure to open a table does not
necessarily prevent execution of a stored routine.
In particular, the pre-locking mechanism is now behaving as follows:
1) the first step, to compute the transitive closure of all the tables
possibly referenced by a statement, is unchanged.
2) the next step, which is to open all the tables involved, only attempts
to open the tables added by the pre-locking code, but silently fails without
reporting any error or invoking any exception handler is the table is not
present. This is achieved by trapping internal errors with
Prelock_error_handler
3) the locking step only locks tables that were successfully opened.
4) when executing sub statements, the list of tables used by each statements
is evaluated as before. The tables needed by the sub statement are expected
to be already opened and locked. Statement referencing tables that were not
opened in step 2) will fail to find the table in the open list, and only at
this point will execution of the user code fail.
5) when a runtime exception is raised at 4), the instruction continuation
destination (the next instruction to execute in case of SQL continue
handlers) is evaluated.
This is achieved with sp_instr::exec_open_and_lock_tables()
6) if a user exception handler is present in the stored routine, that
handler is invoked as usual, so that ER_NO_SUCH_TABLE exceptions can be
trapped by stored routines. If no handler exists, then the runtime execution
will fail as expected.
With all these changes, a side effect is that view security is impacted, in
two different ways.
First, a view defined as "select stored_function()", where the stored
function references a table that may not exist, is considered valid.
The rationale is that, because the stored function might trap exceptions
during execution and still return a valid result, there is no way to decide
when the view is created if a missing table really cause the view to be invalid.
Secondly, testing for existence of tables is now done later during
execution. View security, which consist of trapping errors and return a
generic ER_VIEW_INVALID (to prevent disclosing information) was only
implemented at very specific phases covering *opening* tables, but not
covering the runtime execution. Because of this existing limitation,
errors that were previously trapped and converted into ER_VIEW_INVALID are
not trapped, causing table names to be reported to the user.
This change is exposing an existing problem, which is independent and will
be resolved separately.
2007-03-06 03:42:07 +01:00
|
|
|
if (!table->placeholder())
|
2005-03-04 14:35:28 +01:00
|
|
|
{
|
|
|
|
table->table->query_id= thd->query_id;
|
|
|
|
if (check_lock_and_start_stmt(thd, table->table, table->lock_type))
|
|
|
|
{
|
2009-12-01 15:39:03 +01:00
|
|
|
mysql_unlock_tables(thd, thd->lock);
|
2009-12-01 16:04:32 +01:00
|
|
|
thd->lock= 0;
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2005-03-04 14:35:28 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
Let us mark all tables which don't belong to the statement itself,
|
|
|
|
and was marked as occupied during open_tables() as free for reuse.
|
|
|
|
*/
|
|
|
|
mark_real_tables_as_free_for_reuse(first_not_own);
|
2009-12-03 00:09:22 +01:00
|
|
|
DBUG_PRINT("info",("locked_tables_mode= LTM_PRELOCKED"));
|
2009-12-01 15:39:03 +01:00
|
|
|
thd->locked_tables_mode= LTM_PRELOCKED;
|
2005-03-04 14:35:28 +01:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2001-04-19 19:41:19 +02:00
|
|
|
else
|
|
|
|
{
|
2005-03-04 14:35:28 +01:00
|
|
|
TABLE_LIST *first_not_own= thd->lex->first_not_own_table();
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
/*
|
|
|
|
When open_and_lock_tables() is called for a single table out of
|
|
|
|
a table list, the 'next_global' chain is temporarily broken. We
|
|
|
|
may not find 'first_not_own' before the end of the "list".
|
|
|
|
Look for example at those places where open_n_lock_single_table()
|
|
|
|
is called. That function implements the temporary breaking of
|
|
|
|
a table list for opening a single table.
|
|
|
|
*/
|
|
|
|
for (table= tables;
|
|
|
|
table && table != first_not_own;
|
|
|
|
table= table->next_global)
|
2001-04-19 19:41:19 +02:00
|
|
|
{
|
2009-03-27 12:09:15 +01:00
|
|
|
if (table->placeholder())
|
|
|
|
continue;
|
|
|
|
|
|
|
|
/*
|
|
|
|
In a stored function or trigger we should ensure that we won't change
|
|
|
|
a table that is already used by the calling statement.
|
|
|
|
*/
|
2009-12-01 15:39:03 +01:00
|
|
|
if (thd->locked_tables_mode >= LTM_PRELOCKED &&
|
2009-03-27 12:09:15 +01:00
|
|
|
table->lock_type >= TL_WRITE_ALLOW_WRITE)
|
|
|
|
{
|
|
|
|
for (TABLE* opentab= thd->open_tables; opentab; opentab= opentab->next)
|
|
|
|
{
|
2009-03-27 13:55:14 +01:00
|
|
|
if (table->table->s == opentab->s && opentab->query_id &&
|
|
|
|
table->table->query_id != opentab->query_id)
|
2009-03-27 12:09:15 +01:00
|
|
|
{
|
|
|
|
my_error(ER_CANT_UPDATE_USED_TABLE_IN_SF_OR_TRG, MYF(0),
|
2009-03-27 13:55:14 +01:00
|
|
|
table->table->s->table_name.str);
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2009-03-27 12:09:15 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (check_lock_and_start_stmt(thd, table->table, table->lock_type))
|
2001-04-19 19:41:19 +02:00
|
|
|
{
|
2009-12-09 10:04:55 +01:00
|
|
|
DBUG_RETURN(TRUE);
|
2001-04-19 19:41:19 +02:00
|
|
|
}
|
|
|
|
}
|
2005-03-04 14:35:28 +01:00
|
|
|
/*
|
|
|
|
If we are under explicit LOCK TABLES and our statement requires
|
|
|
|
prelocking, we should mark all "additional" tables as free for use
|
|
|
|
and enter prelocked mode.
|
|
|
|
*/
|
|
|
|
if (thd->lex->requires_prelocking())
|
|
|
|
{
|
|
|
|
mark_real_tables_as_free_for_reuse(first_not_own);
|
2009-12-01 15:39:03 +01:00
|
|
|
DBUG_PRINT("info",
|
|
|
|
("thd->locked_tables_mode= LTM_PRELOCKED_UNDER_LOCK_TABLES"));
|
|
|
|
thd->locked_tables_mode= LTM_PRELOCKED_UNDER_LOCK_TABLES;
|
2005-03-04 14:35:28 +01:00
|
|
|
}
|
2001-04-19 19:41:19 +02:00
|
|
|
}
|
2007-05-29 17:13:17 +02:00
|
|
|
|
|
|
|
DBUG_RETURN(decide_logging_format(thd, tables));
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2002-11-16 19:19:10 +01:00
|
|
|
|
2005-09-15 01:56:09 +02:00
|
|
|
/*
|
|
|
|
Prepare statement for reopening of tables and recalculation of set of
|
|
|
|
prelocked tables.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
close_tables_for_reopen()
|
2006-02-16 14:19:24 +01:00
|
|
|
thd in Thread context
|
|
|
|
tables in/out List of tables which we were trying to open and lock
|
2005-09-15 01:56:09 +02:00
|
|
|
|
|
|
|
*/
|
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
void close_tables_for_reopen(THD *thd, TABLE_LIST **tables)
|
2005-09-15 01:56:09 +02:00
|
|
|
{
|
2009-12-09 10:44:05 +01:00
|
|
|
TABLE_LIST *first_not_own_table= thd->lex->first_not_own_table();
|
|
|
|
TABLE_LIST *tmp;
|
|
|
|
|
2006-02-16 14:19:24 +01:00
|
|
|
/*
|
|
|
|
If table list consists only from tables from prelocking set, table list
|
|
|
|
for new attempt should be empty, so we have to update list's root pointer.
|
|
|
|
*/
|
2009-12-09 10:44:05 +01:00
|
|
|
if (first_not_own_table == *tables)
|
2006-02-16 14:19:24 +01:00
|
|
|
*tables= 0;
|
2005-09-15 01:56:09 +02:00
|
|
|
thd->lex->chop_off_not_own_tables();
|
|
|
|
sp_remove_not_own_routines(thd->lex);
|
2009-12-09 10:44:05 +01:00
|
|
|
for (tmp= *tables; tmp; tmp= tmp->next_global)
|
2009-12-08 10:57:07 +01:00
|
|
|
{
|
2006-02-16 14:19:24 +01:00
|
|
|
tmp->table= 0;
|
2009-12-08 10:57:07 +01:00
|
|
|
tmp->mdl_request.ticket= NULL;
|
|
|
|
}
|
2009-12-09 10:44:05 +01:00
|
|
|
/*
|
|
|
|
Metadata lock requests for tables from extended part of prelocking set
|
|
|
|
are part of list of requests to be waited for in Open_table_context.
|
|
|
|
So to satisfy assumptions in MDL_context::wait_for_locks(), which will
|
|
|
|
performs the waiting, we have to reset MDL_request::ticket values for
|
|
|
|
them as well.
|
|
|
|
*/
|
|
|
|
for (tmp= first_not_own_table; tmp; tmp= tmp->next_global)
|
|
|
|
tmp->mdl_request.ticket= NULL;
|
2009-12-08 10:57:07 +01:00
|
|
|
close_thread_tables(thd);
|
Backport of revno ## 2617.31.1, 2617.31.3, 2617.31.4, 2617.31.5,
2617.31.12, 2617.31.15, 2617.31.15, 2617.31.16, 2617.43.1
- initial changeset that introduced the fix for
Bug#989 and follow up fixes for all test suite failures
introduced in the initial changeset.
------------------------------------------------------------
revno: 2617.31.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Fri 2009-03-06 19:17:00 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
Currently the MySQL server does not keep metadata locks on
schema objects for the duration of a transaction, thus failing
to guarantee the integrity of the schema objects being used
during the transaction and to protect then from concurrent
DDL operations. This also poses a problem for replication as
a DDL operation might be replicated even thought there are
active transactions using the object being modified.
The solution is to defer the release of metadata locks until
a active transaction is either committed or rolled back. This
prevents other statements from modifying the table for the
entire duration of the transaction. This provides commitment
ordering for guaranteeing serializability across multiple
transactions.
- Incompatible change:
If MySQL's metadata locking system encounters a lock conflict,
the usual schema is to use the try and back-off technique to
avoid deadlocks -- this schema consists in releasing all locks
and trying to acquire them all in one go.
But in a transactional context this algorithm can't be utilized
as its not possible to release locks acquired during the course
of the transaction without breaking the transaction commitments.
To avoid deadlocks in this case, the ER_LOCK_DEADLOCK will be
returned if a lock conflict is encountered during a transaction.
Let's consider an example:
A transaction has two statements that modify table t1, then table
t2, and then commits. The first statement of the transaction will
acquire a shared metadata lock on table t1, and it will be kept
utill COMMIT to ensure serializability.
At the moment when the second statement attempts to acquire a
shared metadata lock on t2, a concurrent ALTER or DROP statement
might have locked t2 exclusively. The prescription of the current
locking protocol is that the acquirer of the shared lock backs off
-- gives up all his current locks and retries. This implies that
the entire multi-statement transaction has to be rolled back.
- Incompatible change:
FLUSH commands such as FLUSH PRIVILEGES and FLUSH TABLES WITH READ
LOCK won't cause locked tables to be implicitly unlocked anymore.
2009-12-05 00:02:48 +01:00
|
|
|
if (!thd->locked_tables_mode)
|
|
|
|
thd->mdl_context.release_all_locks();
|
2005-09-15 01:56:09 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
/*
|
2002-11-16 19:19:10 +01:00
|
|
|
Open a single table without table caching and don't set it in open_list
|
2005-11-23 21:45:02 +01:00
|
|
|
|
|
|
|
SYNPOSIS
|
|
|
|
open_temporary_table()
|
|
|
|
thd Thread object
|
|
|
|
path Path (without .frm)
|
|
|
|
db database
|
|
|
|
table_name Table name
|
|
|
|
link_in_list 1 if table should be linked into thd->temporary_tables
|
|
|
|
|
|
|
|
NOTES:
|
|
|
|
Used by alter_table to open a temporary table and when creating
|
|
|
|
a temporary table with CREATE TEMPORARY ...
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
0 Error
|
|
|
|
# TABLE object
|
2000-07-31 21:29:14 +02:00
|
|
|
*/
|
|
|
|
|
|
|
|
TABLE *open_temporary_table(THD *thd, const char *path, const char *db,
|
|
|
|
const char *table_name, bool link_in_list)
|
|
|
|
{
|
|
|
|
TABLE *tmp_table;
|
2005-01-06 12:00:13 +01:00
|
|
|
TABLE_SHARE *share;
|
2005-11-23 21:45:02 +01:00
|
|
|
char cache_key[MAX_DBKEY_LENGTH], *saved_cache_key, *tmp_path;
|
|
|
|
uint key_length;
|
|
|
|
TABLE_LIST table_list;
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("open_temporary_table");
|
2007-02-23 12:13:55 +01:00
|
|
|
DBUG_PRINT("enter",
|
|
|
|
("table: '%s'.'%s' path: '%s' server_id: %u "
|
|
|
|
"pseudo_thread_id: %lu",
|
|
|
|
db, table_name, path,
|
|
|
|
(uint) thd->server_id, (ulong) thd->variables.pseudo_thread_id));
|
2000-12-07 14:48:05 +01:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
table_list.db= (char*) db;
|
|
|
|
table_list.table_name= (char*) table_name;
|
|
|
|
/* Create the cache_key for temporary tables */
|
|
|
|
key_length= create_table_def_key(thd, cache_key, &table_list, 1);
|
|
|
|
|
|
|
|
if (!(tmp_table= (TABLE*) my_malloc(sizeof(*tmp_table) + sizeof(*share) +
|
|
|
|
strlen(path)+1 + key_length,
|
|
|
|
MYF(MY_WME))))
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_RETURN(0); /* purecov: inspected */
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
share= (TABLE_SHARE*) (tmp_table+1);
|
|
|
|
tmp_path= (char*) (share+1);
|
|
|
|
saved_cache_key= strmov(tmp_path, path)+1;
|
|
|
|
memcpy(saved_cache_key, cache_key, key_length);
|
|
|
|
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
init_tmp_table_share(thd, share, saved_cache_key, key_length,
|
2005-11-23 21:45:02 +01:00
|
|
|
strend(saved_cache_key)+1, tmp_path);
|
|
|
|
|
|
|
|
if (open_table_def(thd, share, 0) ||
|
|
|
|
open_table_from_share(thd, share, table_name,
|
|
|
|
(uint) (HA_OPEN_KEYFILE | HA_OPEN_RNDFILE |
|
|
|
|
HA_GET_INDEX),
|
|
|
|
READ_KEYINFO | COMPUTE_TYPES | EXTRA_RECORD,
|
|
|
|
ha_open_options,
|
2006-01-17 08:40:00 +01:00
|
|
|
tmp_table, FALSE))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
/* No need to lock share->mutex as this is not needed for tmp tables */
|
|
|
|
free_table_share(share);
|
2002-10-24 00:29:29 +02:00
|
|
|
my_free((char*) tmp_table,MYF(0));
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
tmp_table->reginfo.lock_type= TL_WRITE; // Simulate locked
|
2005-01-06 12:00:13 +01:00
|
|
|
share->tmp_table= (tmp_table->file->has_transactions() ?
|
2007-03-22 15:07:32 +01:00
|
|
|
TRANSACTIONAL_TMP_TABLE : NON_TRANSACTIONAL_TMP_TABLE);
|
2000-12-07 14:48:05 +01:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
if (link_in_list)
|
|
|
|
{
|
2006-06-19 19:14:14 +02:00
|
|
|
/* growing temp list at the head */
|
|
|
|
tmp_table->next= thd->temporary_tables;
|
|
|
|
if (tmp_table->next)
|
|
|
|
tmp_table->next->prev= tmp_table;
|
2005-11-23 21:45:02 +01:00
|
|
|
thd->temporary_tables= tmp_table;
|
2006-06-19 19:14:14 +02:00
|
|
|
thd->temporary_tables->prev= 0;
|
2001-11-25 12:50:48 +01:00
|
|
|
if (thd->slave_thread)
|
|
|
|
slave_open_temp_tables++;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2006-11-01 02:31:56 +01:00
|
|
|
tmp_table->pos_in_table_list= 0;
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
DBUG_PRINT("tmptable", ("opened table: '%s'.'%s' 0x%lx", tmp_table->s->db.str,
|
|
|
|
tmp_table->s->table_name.str, (long) tmp_table));
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_RETURN(tmp_table);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-12-21 19:18:40 +01:00
|
|
|
bool rm_temporary_table(handlerton *base, char *path)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
bool error=0;
|
2005-11-23 21:45:02 +01:00
|
|
|
handler *file;
|
|
|
|
char *ext;
|
2003-03-19 20:23:13 +01:00
|
|
|
DBUG_ENTER("rm_temporary_table");
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
strmov(ext= strend(path), reg_ext);
|
2000-07-31 21:29:14 +02:00
|
|
|
if (my_delete(path,MYF(0)))
|
|
|
|
error=1; /* purecov: inspected */
|
2005-11-23 21:45:02 +01:00
|
|
|
*ext= 0; // remove extension
|
|
|
|
file= get_new_handler((TABLE_SHARE*) 0, current_thd->mem_root, base);
|
2007-12-20 19:16:55 +01:00
|
|
|
if (file && file->ha_delete_table(path))
|
2004-06-25 15:52:01 +02:00
|
|
|
{
|
2000-07-31 21:29:14 +02:00
|
|
|
error=1;
|
2005-11-23 21:45:02 +01:00
|
|
|
sql_print_warning("Could not remove temporary table: '%s', error: %d",
|
2004-09-04 20:17:09 +02:00
|
|
|
path, my_errno);
|
2004-06-25 15:52:01 +02:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
delete file;
|
2003-03-19 20:23:13 +01:00
|
|
|
DBUG_RETURN(error);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*****************************************************************************
|
2005-08-12 16:57:19 +02:00
|
|
|
* The following find_field_in_XXX procedures implement the core of the
|
|
|
|
* name resolution functionality. The entry point to resolve a column name in a
|
|
|
|
* list of tables is 'find_field_in_tables'. It calls 'find_field_in_table_ref'
|
|
|
|
* for each table reference. In turn, depending on the type of table reference,
|
|
|
|
* 'find_field_in_table_ref' calls one of the 'find_field_in_XXX' procedures
|
|
|
|
* below specific for the type of table reference.
|
2000-07-31 21:29:14 +02:00
|
|
|
******************************************************************************/
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
/* Special Field pointers as return values of find_field_in_XXX functions. */
|
2004-11-12 20:24:16 +01:00
|
|
|
Field *not_found_field= (Field*) 0x1;
|
|
|
|
Field *view_ref_found= (Field*) 0x2;
|
2004-07-16 00:15:55 +02:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
#define WRONG_GRANT (Field*) -1
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
static void update_field_dependencies(THD *thd, Field *field, TABLE *table)
|
|
|
|
{
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
DBUG_ENTER("update_field_dependencies");
|
|
|
|
if (thd->mark_used_columns != MARK_COLUMNS_NONE)
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
MY_BITMAP *current_bitmap, *other_bitmap;
|
|
|
|
|
|
|
|
/*
|
|
|
|
We always want to register the used keys, as the column bitmap may have
|
|
|
|
been set for all fields (for example for view).
|
|
|
|
*/
|
|
|
|
|
2007-03-05 18:08:41 +01:00
|
|
|
table->covering_keys.intersect(field->part_of_key);
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
table->merge_keys.merge(field->part_of_key);
|
|
|
|
|
|
|
|
if (thd->mark_used_columns == MARK_COLUMNS_READ)
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
current_bitmap= table->read_set;
|
|
|
|
other_bitmap= table->write_set;
|
2005-08-12 16:57:19 +02:00
|
|
|
}
|
|
|
|
else
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
{
|
|
|
|
current_bitmap= table->write_set;
|
|
|
|
other_bitmap= table->read_set;
|
|
|
|
}
|
|
|
|
|
2009-05-05 11:38:19 +02:00
|
|
|
/*
|
|
|
|
The test-and-set mechanism in the bitmap is not reliable during
|
|
|
|
multi-UPDATE statements under MARK_COLUMNS_READ mode
|
|
|
|
(thd->mark_used_columns == MARK_COLUMNS_READ), as this bitmap contains
|
|
|
|
only those columns that are used in the SET clause. I.e they are being
|
|
|
|
set here. See multi_update::prepare()
|
|
|
|
*/
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
if (bitmap_fast_test_and_set(current_bitmap, field->field_index))
|
|
|
|
{
|
|
|
|
if (thd->mark_used_columns == MARK_COLUMNS_WRITE)
|
|
|
|
{
|
|
|
|
DBUG_PRINT("warning", ("Found duplicated field"));
|
|
|
|
thd->dup_field= field;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
DBUG_PRINT("note", ("Field found before"));
|
|
|
|
}
|
|
|
|
DBUG_VOID_RETURN;
|
|
|
|
}
|
|
|
|
if (table->get_fields_in_item_tree)
|
|
|
|
field->flags|= GET_FIXED_FIELDS_FLAG;
|
|
|
|
table->used_fields++;
|
|
|
|
}
|
|
|
|
else if (table->get_fields_in_item_tree)
|
2005-08-25 19:50:11 +02:00
|
|
|
field->flags|= GET_FIXED_FIELDS_FLAG;
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
DBUG_VOID_RETURN;
|
2005-08-12 16:57:19 +02:00
|
|
|
}
|
|
|
|
|
2004-07-16 00:15:55 +02:00
|
|
|
|
|
|
|
/*
|
2005-08-12 16:57:19 +02:00
|
|
|
Find a field by name in a view that uses merge algorithm.
|
2004-07-16 00:15:55 +02:00
|
|
|
|
|
|
|
SYNOPSIS
|
2005-08-12 16:57:19 +02:00
|
|
|
find_field_in_view()
|
2004-07-16 00:15:55 +02:00
|
|
|
thd thread handler
|
2005-08-12 16:57:19 +02:00
|
|
|
table_list view to search for 'name'
|
2004-07-16 00:15:55 +02:00
|
|
|
name name of field
|
|
|
|
length length of name
|
2005-11-28 20:57:50 +01:00
|
|
|
item_name name of item if it will be created (VIEW)
|
2005-08-12 16:57:19 +02:00
|
|
|
ref expression substituted in VIEW should be passed
|
|
|
|
using this reference (return view_ref_found)
|
2004-11-21 18:33:49 +01:00
|
|
|
register_tree_change TRUE if ref is not stack variable and we
|
2005-08-12 16:57:19 +02:00
|
|
|
need register changes in item tree
|
2004-07-16 00:15:55 +02:00
|
|
|
|
|
|
|
RETURN
|
|
|
|
0 field is not found
|
|
|
|
view_ref_found found value in VIEW (real result is in *ref)
|
2005-08-12 16:57:19 +02:00
|
|
|
# pointer to field - only for schema table fields
|
2004-07-16 00:15:55 +02:00
|
|
|
*/
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
static Field *
|
|
|
|
find_field_in_view(THD *thd, TABLE_LIST *table_list,
|
2005-11-28 20:57:50 +01:00
|
|
|
const char *name, uint length,
|
|
|
|
const char *item_name, Item **ref,
|
2005-08-12 16:57:19 +02:00
|
|
|
bool register_tree_change)
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_ENTER("find_field_in_view");
|
|
|
|
DBUG_PRINT("enter",
|
|
|
|
("view: '%s', field name: '%s', item name: '%s', ref 0x%lx",
|
|
|
|
table_list->alias, name, item_name, (ulong) ref));
|
|
|
|
Field_iterator_view field_it;
|
|
|
|
field_it.set(table_list);
|
2007-07-28 20:10:38 +02:00
|
|
|
Query_arena *arena= 0, backup;
|
2005-09-08 21:30:05 +02:00
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_ASSERT(table_list->schema_table_reformed ||
|
|
|
|
(ref != 0 && table_list->view != 0));
|
|
|
|
for (; !field_it.end_of_fields(); field_it.next())
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
if (!my_strcasecmp(system_charset_info, field_it.name(), name))
|
2005-01-24 16:44:54 +01:00
|
|
|
{
|
2005-09-08 21:30:05 +02:00
|
|
|
// in PS use own arena or data will be freed after prepare
|
2007-07-28 20:10:38 +02:00
|
|
|
if (register_tree_change && thd->stmt_arena->is_stmt_prepare_or_first_sp_execute())
|
2005-09-08 21:30:05 +02:00
|
|
|
arena= thd->activate_stmt_arena_if_needed(&backup);
|
2005-11-28 20:57:50 +01:00
|
|
|
/*
|
|
|
|
create_item() may, or may not create a new Item, depending on
|
|
|
|
the column reference. See create_view_field() for details.
|
|
|
|
*/
|
2005-08-12 16:57:19 +02:00
|
|
|
Item *item= field_it.create_item(thd);
|
2007-07-28 20:10:38 +02:00
|
|
|
if (arena)
|
2005-09-08 21:30:05 +02:00
|
|
|
thd->restore_active_arena(arena, &backup);
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
if (!item)
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
/*
|
|
|
|
*ref != NULL means that *ref contains the item that we need to
|
|
|
|
replace. If the item was aliased by the user, set the alias to
|
|
|
|
the replacing item.
|
2005-09-12 00:46:35 +02:00
|
|
|
We need to set alias on both ref itself and on ref real item.
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
|
|
|
if (*ref && !(*ref)->is_autogenerated_name)
|
2005-09-12 00:46:35 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
item->set_name((*ref)->name, (*ref)->name_length,
|
|
|
|
system_charset_info);
|
2005-09-12 00:46:35 +02:00
|
|
|
item->real_item()->set_name((*ref)->name, (*ref)->name_length,
|
|
|
|
system_charset_info);
|
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
if (register_tree_change)
|
|
|
|
thd->change_item_tree(ref, item);
|
|
|
|
else
|
|
|
|
*ref= item;
|
|
|
|
DBUG_RETURN((Field*) view_ref_found);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Find field by name in a NATURAL/USING join table reference.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
find_field_in_natural_join()
|
|
|
|
thd [in] thread handler
|
|
|
|
table_ref [in] table reference to search
|
|
|
|
name [in] name of field
|
|
|
|
length [in] length of name
|
|
|
|
ref [in/out] if 'name' is resolved to a view field, ref is
|
|
|
|
set to point to the found view field
|
|
|
|
register_tree_change [in] TRUE if ref is not stack variable and we
|
|
|
|
need register changes in item tree
|
|
|
|
actual_table [out] the original table reference where the field
|
|
|
|
belongs - differs from 'table_list' only for
|
|
|
|
NATURAL/USING joins
|
|
|
|
|
2005-08-23 17:08:04 +02:00
|
|
|
DESCRIPTION
|
|
|
|
Search for a field among the result fields of a NATURAL/USING join.
|
|
|
|
Notice that this procedure is called only for non-qualified field
|
|
|
|
names. In the case of qualified fields, we search directly the base
|
|
|
|
tables of a natural join.
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
RETURN
|
2005-08-17 12:46:36 +02:00
|
|
|
NULL if the field was not found
|
|
|
|
WRONG_GRANT if no access rights to the found field
|
|
|
|
# Pointer to the found Field
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
|
|
|
|
|
|
|
static Field *
|
|
|
|
find_field_in_natural_join(THD *thd, TABLE_LIST *table_ref, const char *name,
|
2005-11-30 20:27:11 +01:00
|
|
|
uint length, Item **ref, bool register_tree_change,
|
2005-08-17 12:46:36 +02:00
|
|
|
TABLE_LIST **actual_table)
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
2005-08-17 12:46:36 +02:00
|
|
|
List_iterator_fast<Natural_join_column>
|
|
|
|
field_it(*(table_ref->join_columns));
|
2007-01-31 15:04:38 +01:00
|
|
|
Natural_join_column *nj_col, *curr_nj_col;
|
2005-08-17 12:46:36 +02:00
|
|
|
Field *found_field;
|
2005-09-08 21:30:05 +02:00
|
|
|
Query_arena *arena, backup;
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_ENTER("find_field_in_natural_join");
|
2005-08-17 12:46:36 +02:00
|
|
|
DBUG_PRINT("enter", ("field name: '%s', ref 0x%lx",
|
2005-08-12 16:57:19 +02:00
|
|
|
name, (ulong) ref));
|
|
|
|
DBUG_ASSERT(table_ref->is_natural_join && table_ref->join_columns);
|
2005-08-17 12:46:36 +02:00
|
|
|
DBUG_ASSERT(*actual_table == NULL);
|
2005-08-12 16:57:19 +02:00
|
|
|
|
2006-03-29 13:27:36 +02:00
|
|
|
LINT_INIT(arena);
|
2005-08-17 12:46:36 +02:00
|
|
|
LINT_INIT(found_field);
|
2005-08-12 16:57:19 +02:00
|
|
|
|
2007-01-31 15:04:38 +01:00
|
|
|
for (nj_col= NULL, curr_nj_col= field_it++; curr_nj_col;
|
|
|
|
curr_nj_col= field_it++)
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
2007-01-31 15:04:38 +01:00
|
|
|
if (!my_strcasecmp(system_charset_info, curr_nj_col->name(), name))
|
|
|
|
{
|
|
|
|
if (nj_col)
|
|
|
|
{
|
|
|
|
my_error(ER_NON_UNIQ_ERROR, MYF(0), name, thd->where);
|
|
|
|
DBUG_RETURN(NULL);
|
|
|
|
}
|
|
|
|
nj_col= curr_nj_col;
|
|
|
|
}
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
2007-01-31 15:04:38 +01:00
|
|
|
if (!nj_col)
|
|
|
|
DBUG_RETURN(NULL);
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
if (nj_col->view_field)
|
|
|
|
{
|
2005-10-11 23:58:22 +02:00
|
|
|
Item *item;
|
2007-02-23 12:13:55 +01:00
|
|
|
LINT_INIT(arena);
|
2005-09-08 21:30:05 +02:00
|
|
|
if (register_tree_change)
|
|
|
|
arena= thd->activate_stmt_arena_if_needed(&backup);
|
2005-11-28 20:57:50 +01:00
|
|
|
/*
|
|
|
|
create_item() may, or may not create a new Item, depending on the
|
|
|
|
column reference. See create_view_field() for details.
|
|
|
|
*/
|
2005-10-11 23:58:22 +02:00
|
|
|
item= nj_col->create_item(thd);
|
2007-01-11 18:10:01 +01:00
|
|
|
/*
|
|
|
|
*ref != NULL means that *ref contains the item that we need to
|
|
|
|
replace. If the item was aliased by the user, set the alias to
|
|
|
|
the replacing item.
|
|
|
|
We need to set alias on both ref itself and on ref real item.
|
|
|
|
*/
|
|
|
|
if (*ref && !(*ref)->is_autogenerated_name)
|
|
|
|
{
|
|
|
|
item->set_name((*ref)->name, (*ref)->name_length,
|
|
|
|
system_charset_info);
|
|
|
|
item->real_item()->set_name((*ref)->name, (*ref)->name_length,
|
|
|
|
system_charset_info);
|
|
|
|
}
|
2005-09-08 21:30:05 +02:00
|
|
|
if (register_tree_change && arena)
|
|
|
|
thd->restore_active_arena(arena, &backup);
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
if (!item)
|
|
|
|
DBUG_RETURN(NULL);
|
|
|
|
DBUG_ASSERT(nj_col->table_field == NULL);
|
|
|
|
if (nj_col->table_ref->schema_table_reformed)
|
2005-08-17 12:46:36 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
/*
|
|
|
|
Translation table items are always Item_fields and fixed
|
|
|
|
already('mysql_schema_table' function). So we can return
|
|
|
|
->field. It is used only for 'show & where' commands.
|
|
|
|
*/
|
|
|
|
DBUG_RETURN(((Item_field*) (nj_col->view_field->item))->field);
|
2005-08-17 12:46:36 +02:00
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
if (register_tree_change)
|
|
|
|
thd->change_item_tree(ref, item);
|
|
|
|
else
|
|
|
|
*ref= item;
|
|
|
|
found_field= (Field*) view_ref_found;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/* This is a base table. */
|
|
|
|
DBUG_ASSERT(nj_col->view_field == NULL);
|
2008-10-07 23:34:00 +02:00
|
|
|
/*
|
|
|
|
This fix_fields is not necessary (initially this item is fixed by
|
|
|
|
the Item_field constructor; after reopen_tables the Item_func_eq
|
|
|
|
calls fix_fields on that item), it's just a check during table
|
|
|
|
reopening for columns that was dropped by the concurrent connection.
|
|
|
|
*/
|
|
|
|
if (!nj_col->table_field->fixed &&
|
|
|
|
nj_col->table_field->fix_fields(thd, (Item **)&nj_col->table_field))
|
|
|
|
{
|
|
|
|
DBUG_PRINT("info", ("column '%s' was dropped by the concurrent connection",
|
|
|
|
nj_col->table_field->name));
|
|
|
|
DBUG_RETURN(NULL);
|
|
|
|
}
|
|
|
|
DBUG_ASSERT(nj_col->table_ref->table == nj_col->table_field->field->table);
|
|
|
|
found_field= nj_col->table_field->field;
|
2005-08-12 16:57:19 +02:00
|
|
|
update_field_dependencies(thd, found_field, nj_col->table_ref->table);
|
|
|
|
}
|
|
|
|
|
|
|
|
*actual_table= nj_col->table_ref;
|
|
|
|
|
|
|
|
DBUG_RETURN(found_field);
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
|
|
|
|
2004-09-03 20:43:04 +02:00
|
|
|
|
2004-07-16 00:15:55 +02:00
|
|
|
/*
|
2005-08-12 16:57:19 +02:00
|
|
|
Find field by name in a base table or a view with temp table algorithm.
|
2004-07-16 00:15:55 +02:00
|
|
|
|
|
|
|
SYNOPSIS
|
2005-08-12 16:57:19 +02:00
|
|
|
find_field_in_table()
|
2004-07-16 00:15:55 +02:00
|
|
|
thd thread handler
|
2005-08-12 16:57:19 +02:00
|
|
|
table table where to search for the field
|
2004-07-16 00:15:55 +02:00
|
|
|
name name of field
|
|
|
|
length length of name
|
|
|
|
allow_rowid do allow finding of "_rowid" field?
|
2005-08-12 16:57:19 +02:00
|
|
|
cached_field_index_ptr cached position in field list (used to speedup
|
|
|
|
lookup for fields in prepared tables)
|
2004-07-16 00:15:55 +02:00
|
|
|
|
|
|
|
RETURN
|
2005-08-17 12:46:36 +02:00
|
|
|
0 field is not found
|
|
|
|
# pointer to field
|
2004-07-16 00:15:55 +02:00
|
|
|
*/
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
Field *
|
|
|
|
find_field_in_table(THD *thd, TABLE *table, const char *name, uint length,
|
2005-11-30 20:27:11 +01:00
|
|
|
bool allow_rowid, uint *cached_field_index_ptr)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2004-06-25 15:52:01 +02:00
|
|
|
Field **field_ptr, *field;
|
|
|
|
uint cached_field_index= *cached_field_index_ptr;
|
2005-08-17 12:46:36 +02:00
|
|
|
DBUG_ENTER("find_field_in_table");
|
|
|
|
DBUG_PRINT("enter", ("table: '%s', field name: '%s'", table->alias, name));
|
2004-06-25 15:52:01 +02:00
|
|
|
|
|
|
|
/* We assume here that table->field < NO_CACHED_FIELD_INDEX = UINT_MAX */
|
2005-01-06 12:00:13 +01:00
|
|
|
if (cached_field_index < table->s->fields &&
|
2005-10-27 23:18:23 +02:00
|
|
|
!my_strcasecmp(system_charset_info,
|
2004-06-25 15:52:01 +02:00
|
|
|
table->field[cached_field_index]->field_name, name))
|
|
|
|
field_ptr= table->field + cached_field_index;
|
2005-01-06 12:00:13 +01:00
|
|
|
else if (table->s->name_hash.records)
|
2005-11-23 21:45:02 +01:00
|
|
|
{
|
2009-10-14 18:37:38 +02:00
|
|
|
field_ptr= (Field**) my_hash_search(&table->s->name_hash, (uchar*) name,
|
|
|
|
length);
|
2005-11-23 21:45:02 +01:00
|
|
|
if (field_ptr)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
field_ptr points to field in TABLE_SHARE. Convert it to the matching
|
|
|
|
field in table
|
|
|
|
*/
|
|
|
|
field_ptr= (table->field + (field_ptr - table->s->field));
|
|
|
|
}
|
|
|
|
}
|
2004-06-25 15:52:01 +02:00
|
|
|
else
|
|
|
|
{
|
|
|
|
if (!(field_ptr= table->field))
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN((Field *)0);
|
2004-06-25 15:52:01 +02:00
|
|
|
for (; *field_ptr; ++field_ptr)
|
|
|
|
if (!my_strcasecmp(system_charset_info, (*field_ptr)->field_name, name))
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (field_ptr && *field_ptr)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2004-06-25 15:52:01 +02:00
|
|
|
*cached_field_index_ptr= field_ptr - table->field;
|
|
|
|
field= *field_ptr;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2004-06-25 15:52:01 +02:00
|
|
|
if (!allow_rowid ||
|
|
|
|
my_strcasecmp(system_charset_info, name, "_rowid") ||
|
2005-11-23 21:45:02 +01:00
|
|
|
table->s->rowid_field_offset == 0)
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN((Field*) 0);
|
2005-11-23 21:45:02 +01:00
|
|
|
field= table->field[table->s->rowid_field_offset-1];
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
update_field_dependencies(thd, field, table);
|
|
|
|
|
|
|
|
DBUG_RETURN(field);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Find field in a table reference.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
find_field_in_table_ref()
|
|
|
|
thd [in] thread handler
|
|
|
|
table_list [in] table reference to search
|
|
|
|
name [in] name of field
|
2005-11-28 20:57:50 +01:00
|
|
|
length [in] field length of name
|
2005-08-12 16:57:19 +02:00
|
|
|
item_name [in] name of item if it will be created (VIEW)
|
|
|
|
db_name [in] optional database name that qualifies the
|
2005-11-28 20:57:50 +01:00
|
|
|
table_name [in] optional table name that qualifies the field
|
2005-08-12 16:57:19 +02:00
|
|
|
ref [in/out] if 'name' is resolved to a view field, ref
|
|
|
|
is set to point to the found view field
|
2005-11-30 20:27:11 +01:00
|
|
|
check_privileges [in] check privileges
|
2005-08-12 16:57:19 +02:00
|
|
|
allow_rowid [in] do allow finding of "_rowid" field?
|
|
|
|
cached_field_index_ptr [in] cached position in field list (used to
|
|
|
|
speedup lookup for fields in prepared tables)
|
|
|
|
register_tree_change [in] TRUE if ref is not stack variable and we
|
|
|
|
need register changes in item tree
|
|
|
|
actual_table [out] the original table reference where the field
|
|
|
|
belongs - differs from 'table_list' only for
|
|
|
|
NATURAL_USING joins.
|
|
|
|
|
2005-09-20 13:00:39 +02:00
|
|
|
DESCRIPTION
|
|
|
|
Find a field in a table reference depending on the type of table
|
|
|
|
reference. There are three types of table references with respect
|
|
|
|
to the representation of their result columns:
|
|
|
|
- an array of Field_translator objects for MERGE views and some
|
|
|
|
information_schema tables,
|
|
|
|
- an array of Field objects (and possibly a name hash) for stored
|
|
|
|
tables,
|
|
|
|
- a list of Natural_join_column objects for NATURAL/USING joins.
|
|
|
|
This procedure detects the type of the table reference 'table_list'
|
|
|
|
and calls the corresponding search routine.
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
RETURN
|
|
|
|
0 field is not found
|
|
|
|
view_ref_found found value in VIEW (real result is in *ref)
|
|
|
|
# pointer to field
|
|
|
|
*/
|
|
|
|
|
|
|
|
Field *
|
|
|
|
find_field_in_table_ref(THD *thd, TABLE_LIST *table_list,
|
2005-11-28 20:57:50 +01:00
|
|
|
const char *name, uint length,
|
|
|
|
const char *item_name, const char *db_name,
|
|
|
|
const char *table_name, Item **ref,
|
2005-11-30 20:27:11 +01:00
|
|
|
bool check_privileges, bool allow_rowid,
|
|
|
|
uint *cached_field_index_ptr,
|
2005-08-12 16:57:19 +02:00
|
|
|
bool register_tree_change, TABLE_LIST **actual_table)
|
|
|
|
{
|
|
|
|
Field *fld;
|
|
|
|
DBUG_ENTER("find_field_in_table_ref");
|
2007-02-23 12:13:55 +01:00
|
|
|
DBUG_ASSERT(table_list->alias);
|
|
|
|
DBUG_ASSERT(name);
|
|
|
|
DBUG_ASSERT(item_name);
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_PRINT("enter",
|
|
|
|
("table: '%s' field name: '%s' item name: '%s' ref 0x%lx",
|
|
|
|
table_list->alias, name, item_name, (ulong) ref));
|
|
|
|
|
|
|
|
/*
|
|
|
|
Check that the table and database that qualify the current field name
|
2005-09-20 13:00:39 +02:00
|
|
|
are the same as the table reference we are going to search for the field.
|
|
|
|
|
2005-09-30 09:37:58 +02:00
|
|
|
Exclude from the test below nested joins because the columns in a
|
|
|
|
nested join generally originate from different tables. Nested joins
|
|
|
|
also have no table name, except when a nested join is a merge view
|
|
|
|
or an information schema table.
|
2005-09-20 13:00:39 +02:00
|
|
|
|
|
|
|
We include explicitly table references with a 'field_translation' table,
|
|
|
|
because if there are views over natural joins we don't want to search
|
|
|
|
inside the view, but we want to search directly in the view columns
|
|
|
|
which are represented as a 'field_translation'.
|
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
TODO: Ensure that table_name, db_name and tables->db always points to
|
|
|
|
something !
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
2005-09-29 16:34:21 +02:00
|
|
|
if (/* Exclude nested joins. */
|
|
|
|
(!table_list->nested_join ||
|
2005-09-20 13:00:39 +02:00
|
|
|
/* Include merge views and information schema tables. */
|
|
|
|
table_list->field_translation) &&
|
|
|
|
/*
|
|
|
|
Test if the field qualifiers match the table reference we plan
|
|
|
|
to search.
|
|
|
|
*/
|
2005-08-17 12:46:36 +02:00
|
|
|
table_name && table_name[0] &&
|
|
|
|
(my_strcasecmp(table_alias_charset, table_list->alias, table_name) ||
|
2005-08-12 16:57:19 +02:00
|
|
|
(db_name && db_name[0] && table_list->db && table_list->db[0] &&
|
2009-10-23 11:19:54 +02:00
|
|
|
(table_list->schema_table ?
|
|
|
|
my_strcasecmp(system_charset_info, db_name, table_list->db) :
|
|
|
|
strcmp(db_name, table_list->db)))))
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN(0);
|
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
*actual_table= NULL;
|
2005-09-20 13:00:39 +02:00
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
if (table_list->field_translation)
|
|
|
|
{
|
2005-09-20 13:00:39 +02:00
|
|
|
/* 'table_list' is a view or an information schema table. */
|
2005-11-30 20:27:11 +01:00
|
|
|
if ((fld= find_field_in_view(thd, table_list, name, length, item_name, ref,
|
2005-08-17 12:46:36 +02:00
|
|
|
register_tree_change)))
|
2005-08-12 16:57:19 +02:00
|
|
|
*actual_table= table_list;
|
|
|
|
}
|
2005-09-29 16:34:21 +02:00
|
|
|
else if (!table_list->nested_join)
|
2005-08-23 17:08:04 +02:00
|
|
|
{
|
2005-09-29 16:34:21 +02:00
|
|
|
/* 'table_list' is a stored table. */
|
|
|
|
DBUG_ASSERT(table_list->table);
|
2005-09-20 13:00:39 +02:00
|
|
|
if ((fld= find_field_in_table(thd, table_list->table, name, length,
|
2005-11-30 20:27:11 +01:00
|
|
|
allow_rowid,
|
|
|
|
cached_field_index_ptr)))
|
2005-09-20 13:00:39 +02:00
|
|
|
*actual_table= table_list;
|
|
|
|
}
|
|
|
|
else
|
2005-08-23 17:08:04 +02:00
|
|
|
{
|
2005-09-08 10:29:52 +02:00
|
|
|
/*
|
2005-09-20 13:00:39 +02:00
|
|
|
'table_list' is a NATURAL/USING join, or an operand of such join that
|
|
|
|
is a nested join itself.
|
|
|
|
|
|
|
|
If the field name we search for is qualified, then search for the field
|
|
|
|
in the table references used by NATURAL/USING the join.
|
2005-09-08 10:29:52 +02:00
|
|
|
*/
|
2005-08-23 17:08:04 +02:00
|
|
|
if (table_name && table_name[0])
|
|
|
|
{
|
|
|
|
List_iterator<TABLE_LIST> it(table_list->nested_join->join_list);
|
|
|
|
TABLE_LIST *table;
|
|
|
|
while ((table= it++))
|
|
|
|
{
|
2005-11-28 20:57:50 +01:00
|
|
|
if ((fld= find_field_in_table_ref(thd, table, name, length, item_name,
|
|
|
|
db_name, table_name, ref,
|
2005-11-30 20:27:11 +01:00
|
|
|
check_privileges, allow_rowid,
|
|
|
|
cached_field_index_ptr,
|
2005-08-23 17:08:04 +02:00
|
|
|
register_tree_change, actual_table)))
|
|
|
|
DBUG_RETURN(fld);
|
|
|
|
}
|
|
|
|
DBUG_RETURN(0);
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
Non-qualified field, search directly in the result columns of the
|
2005-09-08 10:29:52 +02:00
|
|
|
natural join. The condition of the outer IF is true for the top-most
|
|
|
|
natural join, thus if the field is not qualified, we will search
|
|
|
|
directly the top-most NATURAL/USING join.
|
2005-08-23 17:08:04 +02:00
|
|
|
*/
|
|
|
|
fld= find_field_in_natural_join(thd, table_list, name, length, ref,
|
2005-08-12 16:57:19 +02:00
|
|
|
register_tree_change, actual_table);
|
2005-08-23 17:08:04 +02:00
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
|
2006-02-08 18:08:18 +01:00
|
|
|
if (fld)
|
|
|
|
{
|
2005-11-30 20:27:11 +01:00
|
|
|
#ifndef NO_EMBEDDED_ACCESS_CHECKS
|
2006-02-08 18:08:18 +01:00
|
|
|
/* Check if there are sufficient access rights to the found field. */
|
|
|
|
if (check_privileges &&
|
|
|
|
check_column_grant_in_table_ref(thd, *actual_table, name, length))
|
|
|
|
fld= WRONG_GRANT;
|
|
|
|
else
|
2005-11-30 20:27:11 +01:00
|
|
|
#endif
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
if (thd->mark_used_columns != MARK_COLUMNS_NONE)
|
2006-02-08 18:08:18 +01:00
|
|
|
{
|
2006-02-08 22:13:47 +01:00
|
|
|
/*
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
Get rw_set correct for this field so that the handler
|
|
|
|
knows that this field is involved in the query and gets
|
|
|
|
retrieved/updated
|
2006-02-08 22:13:47 +01:00
|
|
|
*/
|
2006-02-08 18:08:18 +01:00
|
|
|
Field *field_to_set= NULL;
|
|
|
|
if (fld == view_ref_found)
|
|
|
|
{
|
|
|
|
Item *it= (*ref)->real_item();
|
|
|
|
if (it->type() == Item::FIELD_ITEM)
|
|
|
|
field_to_set= ((Item_field*)it)->field;
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
else
|
|
|
|
{
|
|
|
|
if (thd->mark_used_columns == MARK_COLUMNS_READ)
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
it->walk(&Item::register_field_in_read_map, 1, (uchar *) 0);
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
}
|
2006-02-08 18:08:18 +01:00
|
|
|
}
|
|
|
|
else
|
|
|
|
field_to_set= fld;
|
|
|
|
if (field_to_set)
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
{
|
|
|
|
TABLE *table= field_to_set->table;
|
|
|
|
if (thd->mark_used_columns == MARK_COLUMNS_READ)
|
|
|
|
bitmap_set_bit(table->read_set, field_to_set->field_index);
|
|
|
|
else
|
|
|
|
bitmap_set_bit(table->write_set, field_to_set->field_index);
|
|
|
|
}
|
2006-02-08 18:08:18 +01:00
|
|
|
}
|
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN(fld);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2003-02-12 20:55:37 +01:00
|
|
|
|
2005-08-25 19:50:11 +02:00
|
|
|
/*
|
|
|
|
Find field in table, no side effects, only purpose is to check for field
|
|
|
|
in table object and get reference to the field if found.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
find_field_in_table_sef()
|
|
|
|
|
|
|
|
table table where to find
|
|
|
|
name Name of field searched for
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
0 field is not found
|
|
|
|
# pointer to field
|
|
|
|
*/
|
|
|
|
|
|
|
|
Field *find_field_in_table_sef(TABLE *table, const char *name)
|
|
|
|
{
|
|
|
|
Field **field_ptr;
|
|
|
|
if (table->s->name_hash.records)
|
2006-02-13 20:34:36 +01:00
|
|
|
{
|
2009-10-14 18:37:38 +02:00
|
|
|
field_ptr= (Field**)my_hash_search(&table->s->name_hash,(uchar*) name,
|
|
|
|
strlen(name));
|
2006-02-13 20:34:36 +01:00
|
|
|
if (field_ptr)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
field_ptr points to field in TABLE_SHARE. Convert it to the matching
|
|
|
|
field in table
|
|
|
|
*/
|
|
|
|
field_ptr= (table->field + (field_ptr - table->s->field));
|
|
|
|
}
|
|
|
|
}
|
2005-08-25 19:50:11 +02:00
|
|
|
else
|
|
|
|
{
|
|
|
|
if (!(field_ptr= table->field))
|
|
|
|
return (Field *)0;
|
|
|
|
for (; *field_ptr; ++field_ptr)
|
|
|
|
if (!my_strcasecmp(system_charset_info, (*field_ptr)->field_name, name))
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (field_ptr)
|
|
|
|
return *field_ptr;
|
|
|
|
else
|
|
|
|
return (Field *)0;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2002-10-08 13:50:12 +02:00
|
|
|
/*
|
|
|
|
Find field in table list.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
find_field_in_tables()
|
2005-08-12 16:57:19 +02:00
|
|
|
thd pointer to current thread structure
|
|
|
|
item field item that should be found
|
|
|
|
first_table list of tables to be searched for item
|
|
|
|
last_table end of the list of tables to search for item. If NULL
|
|
|
|
then search to the end of the list 'first_table'.
|
|
|
|
ref if 'item' is resolved to a view field, ref is set to
|
2005-07-02 23:51:02 +02:00
|
|
|
point to the found view field
|
2005-08-12 16:57:19 +02:00
|
|
|
report_error Degree of error reporting:
|
2005-07-02 23:51:02 +02:00
|
|
|
- IGNORE_ERRORS then do not report any error
|
|
|
|
- IGNORE_EXCEPT_NON_UNIQUE report only non-unique
|
2005-08-12 16:57:19 +02:00
|
|
|
fields, suppress all other errors
|
2005-07-02 23:51:02 +02:00
|
|
|
- REPORT_EXCEPT_NON_UNIQUE report all other errors
|
|
|
|
except when non-unique fields were found
|
|
|
|
- REPORT_ALL_ERRORS
|
|
|
|
check_privileges need to check privileges
|
2005-08-12 16:57:19 +02:00
|
|
|
register_tree_change TRUE if ref is not a stack variable and we
|
|
|
|
to need register changes in item tree
|
2002-10-08 13:50:12 +02:00
|
|
|
|
|
|
|
RETURN VALUES
|
2004-11-05 14:48:44 +01:00
|
|
|
0 If error: the found field is not unique, or there are
|
|
|
|
no sufficient access priviliges for the found field,
|
|
|
|
or the field is qualified with non-existing table.
|
2004-11-02 17:23:15 +01:00
|
|
|
not_found_field The function was called with report_error ==
|
|
|
|
(IGNORE_ERRORS || IGNORE_EXCEPT_NON_UNIQUE) and a
|
|
|
|
field was not found.
|
|
|
|
view_ref_found View field is found, item passed through ref parameter
|
|
|
|
found field If a item was resolved to some field
|
2002-10-08 13:50:12 +02:00
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
Field *
|
2005-08-12 16:57:19 +02:00
|
|
|
find_field_in_tables(THD *thd, Item_ident *item,
|
|
|
|
TABLE_LIST *first_table, TABLE_LIST *last_table,
|
2004-11-02 17:23:15 +01:00
|
|
|
Item **ref, find_item_error_report_type report_error,
|
2005-07-02 23:51:02 +02:00
|
|
|
bool check_privileges, bool register_tree_change)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
Field *found=0;
|
2005-08-12 16:57:19 +02:00
|
|
|
const char *db= item->db_name;
|
|
|
|
const char *table_name= item->table_name;
|
|
|
|
const char *name= item->field_name;
|
2000-08-21 23:18:32 +02:00
|
|
|
uint length=(uint) strlen(name);
|
2003-12-14 05:39:52 +01:00
|
|
|
char name_buff[NAME_LEN+1];
|
2005-08-12 16:57:19 +02:00
|
|
|
TABLE_LIST *cur_table= first_table;
|
|
|
|
TABLE_LIST *actual_table;
|
2005-08-24 23:21:47 +02:00
|
|
|
bool allow_rowid;
|
|
|
|
|
|
|
|
if (!table_name || !table_name[0])
|
|
|
|
{
|
|
|
|
table_name= 0; // For easier test
|
|
|
|
db= 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
allow_rowid= table_name || (cur_table && !cur_table->next_local);
|
2005-07-04 02:24:25 +02:00
|
|
|
|
2004-06-25 15:52:01 +02:00
|
|
|
if (item->cached_table)
|
|
|
|
{
|
|
|
|
/*
|
2005-08-12 16:57:19 +02:00
|
|
|
This shortcut is used by prepared statements. We assume that
|
|
|
|
TABLE_LIST *first_table is not changed during query execution (which
|
2005-01-03 23:22:22 +01:00
|
|
|
is true for all queries except RENAME but luckily RENAME doesn't
|
2004-06-25 15:52:01 +02:00
|
|
|
use fields...) so we can rely on reusing pointer to its member.
|
2004-11-02 19:11:00 +01:00
|
|
|
With this optimization we also miss case when addition of one more
|
2005-01-03 23:22:22 +01:00
|
|
|
field makes some prepared query ambiguous and so erroneous, but we
|
2004-06-25 15:52:01 +02:00
|
|
|
accept this trade off.
|
|
|
|
*/
|
2005-08-12 16:57:19 +02:00
|
|
|
TABLE_LIST *table_ref= item->cached_table;
|
|
|
|
/*
|
|
|
|
The condition (table_ref->view == NULL) ensures that we will call
|
|
|
|
find_field_in_table even in the case of information schema tables
|
|
|
|
when table_ref->field_translation != NULL.
|
|
|
|
*/
|
|
|
|
if (table_ref->table && !table_ref->view)
|
|
|
|
found= find_field_in_table(thd, table_ref->table, name, length,
|
2005-11-30 20:27:11 +01:00
|
|
|
TRUE, &(item->cached_field_index));
|
2004-09-14 18:28:29 +02:00
|
|
|
else
|
2005-11-28 20:57:50 +01:00
|
|
|
found= find_field_in_table_ref(thd, table_ref, name, length, item->name,
|
2005-11-30 20:27:11 +01:00
|
|
|
NULL, NULL, ref, check_privileges,
|
|
|
|
TRUE, &(item->cached_field_index),
|
2005-08-12 16:57:19 +02:00
|
|
|
register_tree_change,
|
|
|
|
&actual_table);
|
2004-06-25 15:52:01 +02:00
|
|
|
if (found)
|
|
|
|
{
|
|
|
|
if (found == WRONG_GRANT)
|
2004-09-14 18:28:29 +02:00
|
|
|
return (Field*) 0;
|
2006-11-29 12:52:11 +01:00
|
|
|
|
|
|
|
/*
|
|
|
|
Only views fields should be marked as dependent, not an underlying
|
|
|
|
fields.
|
|
|
|
*/
|
|
|
|
if (!table_ref->belong_to_view)
|
2005-04-29 01:43:56 +02:00
|
|
|
{
|
|
|
|
SELECT_LEX *current_sel= thd->lex->current_select;
|
2005-08-12 16:57:19 +02:00
|
|
|
SELECT_LEX *last_select= table_ref->select_lex;
|
2005-04-30 18:27:22 +02:00
|
|
|
/*
|
|
|
|
If the field was an outer referencee, mark all selects using this
|
2005-08-12 16:57:19 +02:00
|
|
|
sub query as dependent on the outer query
|
2005-04-30 18:27:22 +02:00
|
|
|
*/
|
2005-04-29 01:43:56 +02:00
|
|
|
if (current_sel != last_select)
|
2005-04-30 18:27:22 +02:00
|
|
|
mark_select_range_as_dependent(thd, last_select, current_sel,
|
2005-04-29 01:43:56 +02:00
|
|
|
found, *ref, item);
|
|
|
|
}
|
2004-06-25 15:52:01 +02:00
|
|
|
return found;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2003-12-14 05:39:52 +01:00
|
|
|
if (db && lower_case_table_names)
|
|
|
|
{
|
|
|
|
/*
|
2004-11-02 19:11:00 +01:00
|
|
|
convert database to lower case for comparison.
|
2003-12-14 05:39:52 +01:00
|
|
|
We can't do this in Item_field as this would change the
|
|
|
|
'name' of the item which may be used in the select list
|
|
|
|
*/
|
|
|
|
strmake(name_buff, db, sizeof(name_buff)-1);
|
2003-12-19 15:25:50 +01:00
|
|
|
my_casedn_str(files_charset_info, name_buff);
|
2003-12-14 05:39:52 +01:00
|
|
|
db= name_buff;
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
if (last_table)
|
|
|
|
last_table= last_table->next_name_resolution_table;
|
|
|
|
|
|
|
|
for (; cur_table != last_table ;
|
|
|
|
cur_table= cur_table->next_name_resolution_table)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-11-28 20:57:50 +01:00
|
|
|
Field *cur_field= find_field_in_table_ref(thd, cur_table, name, length,
|
|
|
|
item->name, db, table_name, ref,
|
2007-10-26 09:01:29 +02:00
|
|
|
(thd->lex->sql_command ==
|
|
|
|
SQLCOM_SHOW_FIELDS)
|
|
|
|
? false : check_privileges,
|
2007-03-23 19:24:03 +01:00
|
|
|
allow_rowid,
|
2005-08-12 16:57:19 +02:00
|
|
|
&(item->cached_field_index),
|
|
|
|
register_tree_change,
|
|
|
|
&actual_table);
|
|
|
|
if (cur_field)
|
2002-12-17 19:15:15 +01:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
if (cur_field == WRONG_GRANT)
|
2007-03-23 19:24:03 +01:00
|
|
|
{
|
|
|
|
if (thd->lex->sql_command != SQLCOM_SHOW_FIELDS)
|
|
|
|
return (Field*) 0;
|
|
|
|
|
|
|
|
thd->clear_error();
|
|
|
|
cur_field= find_field_in_table_ref(thd, cur_table, name, length,
|
|
|
|
item->name, db, table_name, ref,
|
|
|
|
false,
|
|
|
|
allow_rowid,
|
|
|
|
&(item->cached_field_index),
|
|
|
|
register_tree_change,
|
|
|
|
&actual_table);
|
|
|
|
if (cur_field)
|
|
|
|
{
|
|
|
|
Field *nf=new Field_null(NULL,0,Field::NONE,
|
|
|
|
cur_field->field_name,
|
|
|
|
&my_charset_bin);
|
2007-03-24 09:59:15 +01:00
|
|
|
nf->init(cur_table->table);
|
2007-03-23 19:24:03 +01:00
|
|
|
cur_field= nf;
|
|
|
|
}
|
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
Store the original table of the field, which may be different from
|
|
|
|
cur_table in the case of NATURAL/USING join.
|
|
|
|
*/
|
|
|
|
item->cached_table= (!actual_table->cacheable_table || found) ?
|
|
|
|
0 : actual_table;
|
|
|
|
|
2005-08-23 17:08:04 +02:00
|
|
|
DBUG_ASSERT(thd->where);
|
|
|
|
/*
|
|
|
|
If we found a fully qualified field we return it directly as it can't
|
|
|
|
have duplicates.
|
|
|
|
*/
|
2005-08-24 23:21:47 +02:00
|
|
|
if (db)
|
2005-08-23 17:08:04 +02:00
|
|
|
return cur_field;
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
if (found)
|
|
|
|
{
|
2004-11-02 17:23:15 +01:00
|
|
|
if (report_error == REPORT_ALL_ERRORS ||
|
|
|
|
report_error == IGNORE_EXCEPT_NON_UNIQUE)
|
2005-08-23 17:08:04 +02:00
|
|
|
my_error(ER_NON_UNIQ_ERROR, MYF(0),
|
2005-08-24 23:21:47 +02:00
|
|
|
table_name ? item->full_name() : name, thd->where);
|
2007-03-23 19:24:03 +01:00
|
|
|
return (Field*) 0;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
found= cur_field;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
}
|
2005-08-23 17:08:04 +02:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
if (found)
|
|
|
|
return found;
|
2005-08-23 17:08:04 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
If the field was qualified and there were no tables to search, issue
|
|
|
|
an error that an unknown table was given. The situation is detected
|
|
|
|
as follows: if there were no tables we wouldn't go through the loop
|
|
|
|
and cur_table wouldn't be updated by the loop increment part, so it
|
|
|
|
will be equal to the first table.
|
|
|
|
*/
|
2005-08-24 23:21:47 +02:00
|
|
|
if (table_name && (cur_table == first_table) &&
|
2005-08-23 17:08:04 +02:00
|
|
|
(report_error == REPORT_ALL_ERRORS ||
|
|
|
|
report_error == REPORT_EXCEPT_NON_UNIQUE))
|
|
|
|
{
|
2009-06-19 10:24:43 +02:00
|
|
|
char buff[NAME_LEN*2 + 2];
|
2005-08-23 17:08:04 +02:00
|
|
|
if (db && db[0])
|
|
|
|
{
|
|
|
|
strxnmov(buff,sizeof(buff)-1,db,".",table_name,NullS);
|
|
|
|
table_name=buff;
|
|
|
|
}
|
|
|
|
my_error(ER_UNKNOWN_TABLE, MYF(0), table_name, thd->where);
|
|
|
|
}
|
2002-10-08 13:50:12 +02:00
|
|
|
else
|
2005-08-24 23:21:47 +02:00
|
|
|
{
|
2005-08-23 17:08:04 +02:00
|
|
|
if (report_error == REPORT_ALL_ERRORS ||
|
|
|
|
report_error == REPORT_EXCEPT_NON_UNIQUE)
|
|
|
|
my_error(ER_BAD_FIELD_ERROR, MYF(0), item->full_name(), thd->where);
|
|
|
|
else
|
|
|
|
found= not_found_field;
|
2005-08-24 23:21:47 +02:00
|
|
|
}
|
2005-08-23 17:08:04 +02:00
|
|
|
return found;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2002-10-08 13:50:12 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
Find Item in list of items (find_field_in_tables analog)
|
2003-01-25 01:25:52 +01:00
|
|
|
|
|
|
|
TODO
|
|
|
|
is it better return only counter?
|
|
|
|
|
2002-10-08 13:50:12 +02:00
|
|
|
SYNOPSIS
|
|
|
|
find_item_in_list()
|
2003-02-12 20:55:37 +01:00
|
|
|
find Item to find
|
|
|
|
items List of items
|
|
|
|
counter To return number of found item
|
2002-10-08 13:50:12 +02:00
|
|
|
report_error
|
2003-02-12 20:55:37 +01:00
|
|
|
REPORT_ALL_ERRORS report errors, return 0 if error
|
|
|
|
REPORT_EXCEPT_NOT_FOUND Do not report 'not found' error and
|
|
|
|
return not_found_item, report other errors,
|
|
|
|
return 0
|
|
|
|
IGNORE_ERRORS Do not report errors, return 0 if error
|
2007-03-05 04:54:35 +01:00
|
|
|
resolution Set to the resolution type if the item is found
|
|
|
|
(it says whether the item is resolved
|
|
|
|
against an alias name,
|
|
|
|
or as a field name without alias,
|
|
|
|
or as a field hidden by alias,
|
|
|
|
or ignoring alias)
|
|
|
|
|
2002-10-08 13:50:12 +02:00
|
|
|
RETURN VALUES
|
2003-02-12 20:55:37 +01:00
|
|
|
0 Item is not found or item is not unique,
|
|
|
|
error message is reported
|
|
|
|
not_found_item Function was called with
|
|
|
|
report_error == REPORT_EXCEPT_NOT_FOUND and
|
|
|
|
item was not found. No error message was reported
|
2004-09-30 14:28:17 +02:00
|
|
|
found field
|
2002-10-08 13:50:12 +02:00
|
|
|
*/
|
|
|
|
|
2004-11-02 17:23:15 +01:00
|
|
|
/* Special Item pointer to serve as a return value from find_item_in_list(). */
|
2004-11-12 20:24:16 +01:00
|
|
|
Item **not_found_item= (Item**) 0x1;
|
2003-02-12 20:55:37 +01:00
|
|
|
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
Item **
|
2003-01-25 01:25:52 +01:00
|
|
|
find_item_in_list(Item *find, List<Item> &items, uint *counter,
|
2007-03-05 04:54:35 +01:00
|
|
|
find_item_error_report_type report_error,
|
|
|
|
enum_resolution_type *resolution)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
List_iterator<Item> li(items);
|
2004-08-20 15:53:46 +02:00
|
|
|
Item **found=0, **found_unaliased= 0, *item;
|
2003-06-04 17:28:51 +02:00
|
|
|
const char *db_name=0;
|
2000-07-31 21:29:14 +02:00
|
|
|
const char *field_name=0;
|
|
|
|
const char *table_name=0;
|
2004-08-20 15:53:46 +02:00
|
|
|
bool found_unaliased_non_uniq= 0;
|
2006-10-16 12:10:25 +02:00
|
|
|
/*
|
|
|
|
true if the item that we search for is a valid name reference
|
|
|
|
(and not an item that happens to have a name).
|
|
|
|
*/
|
|
|
|
bool is_ref_by_name= 0;
|
2009-06-17 16:56:44 +02:00
|
|
|
uint unaliased_counter= 0;
|
2004-12-18 04:19:21 +01:00
|
|
|
|
2007-03-05 04:54:35 +01:00
|
|
|
*resolution= NOT_RESOLVED;
|
2004-09-30 14:28:17 +02:00
|
|
|
|
2006-10-16 12:10:25 +02:00
|
|
|
is_ref_by_name= (find->type() == Item::FIELD_ITEM ||
|
|
|
|
find->type() == Item::REF_ITEM);
|
|
|
|
if (is_ref_by_name)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
field_name= ((Item_ident*) find)->field_name;
|
|
|
|
table_name= ((Item_ident*) find)->table_name;
|
2003-06-04 17:28:51 +02:00
|
|
|
db_name= ((Item_ident*) find)->db_name;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2003-01-30 17:07:39 +01:00
|
|
|
for (uint i= 0; (item=li++); i++)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-10-01 08:35:30 +02:00
|
|
|
if (field_name && item->real_item()->type() == Item::FIELD_ITEM)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-10-01 08:35:30 +02:00
|
|
|
Item_ident *item_field= (Item_ident*) item;
|
2004-08-20 15:53:46 +02:00
|
|
|
|
2004-06-25 15:52:01 +02:00
|
|
|
/*
|
|
|
|
In case of group_concat() with ORDER BY condition in the QUERY
|
|
|
|
item_field can be field of temporary table without item name
|
|
|
|
(if this field created from expression argument of group_concat()),
|
|
|
|
=> we have to check presence of name before compare
|
|
|
|
*/
|
2004-08-20 15:53:46 +02:00
|
|
|
if (!item_field->name)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (table_name)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2004-08-20 15:53:46 +02:00
|
|
|
/*
|
|
|
|
If table name is specified we should find field 'field_name' in
|
|
|
|
table 'table_name'. According to SQL-standard we should ignore
|
2004-09-30 14:28:17 +02:00
|
|
|
aliases in this case.
|
|
|
|
|
|
|
|
Since we should NOT prefer fields from the select list over
|
|
|
|
other fields from the tables participating in this select in
|
|
|
|
case of ambiguity we have to do extra check outside this function.
|
2004-08-20 15:53:46 +02:00
|
|
|
|
2004-08-31 13:35:04 +02:00
|
|
|
We use strcmp for table names and database names as these may be
|
2004-09-30 14:28:17 +02:00
|
|
|
case sensitive. In cases where they are not case sensitive, they
|
|
|
|
are always in lower case.
|
2004-09-06 12:00:24 +02:00
|
|
|
|
|
|
|
item_field->field_name and item_field->table_name can be 0x0 if
|
2004-09-30 14:28:17 +02:00
|
|
|
item is not fix_field()'ed yet.
|
2004-08-20 15:53:46 +02:00
|
|
|
*/
|
2004-09-06 12:00:24 +02:00
|
|
|
if (item_field->field_name && item_field->table_name &&
|
|
|
|
!my_strcasecmp(system_charset_info, item_field->field_name,
|
2004-08-20 15:53:46 +02:00
|
|
|
field_name) &&
|
2007-11-13 10:39:52 +01:00
|
|
|
!my_strcasecmp(table_alias_charset, item_field->table_name,
|
|
|
|
table_name) &&
|
2004-08-20 15:53:46 +02:00
|
|
|
(!db_name || (item_field->db_name &&
|
|
|
|
!strcmp(item_field->db_name, db_name))))
|
|
|
|
{
|
2004-09-30 14:28:17 +02:00
|
|
|
if (found_unaliased)
|
2004-08-20 15:53:46 +02:00
|
|
|
{
|
2004-09-30 14:28:17 +02:00
|
|
|
if ((*found_unaliased)->eq(item, 0))
|
|
|
|
continue;
|
|
|
|
/*
|
|
|
|
Two matching fields in select list.
|
|
|
|
We already can bail out because we are searching through
|
|
|
|
unaliased names only and will have duplicate error anyway.
|
|
|
|
*/
|
2004-08-20 15:53:46 +02:00
|
|
|
if (report_error != IGNORE_ERRORS)
|
2004-11-13 18:35:51 +01:00
|
|
|
my_error(ER_NON_UNIQ_ERROR, MYF(0),
|
|
|
|
find->full_name(), current_thd->where);
|
2004-08-20 15:53:46 +02:00
|
|
|
return (Item**) 0;
|
|
|
|
}
|
2004-09-30 14:28:17 +02:00
|
|
|
found_unaliased= li.ref();
|
|
|
|
unaliased_counter= i;
|
2007-03-05 04:54:35 +01:00
|
|
|
*resolution= RESOLVED_IGNORING_ALIAS;
|
2004-08-31 13:35:04 +02:00
|
|
|
if (db_name)
|
|
|
|
break; // Perfect match
|
2004-08-20 15:53:46 +02:00
|
|
|
}
|
|
|
|
}
|
2007-03-05 04:54:35 +01:00
|
|
|
else
|
2004-08-20 15:53:46 +02:00
|
|
|
{
|
2007-03-05 04:54:35 +01:00
|
|
|
int fname_cmp= my_strcasecmp(system_charset_info,
|
|
|
|
item_field->field_name,
|
|
|
|
field_name);
|
|
|
|
if (!my_strcasecmp(system_charset_info,
|
|
|
|
item_field->name,field_name))
|
2004-08-20 15:53:46 +02:00
|
|
|
{
|
2007-03-05 04:54:35 +01:00
|
|
|
/*
|
|
|
|
If table name was not given we should scan through aliases
|
|
|
|
and non-aliased fields first. We are also checking unaliased
|
|
|
|
name of the field in then next else-if, to be able to find
|
|
|
|
instantly field (hidden by alias) if no suitable alias or
|
|
|
|
non-aliased field was found.
|
|
|
|
*/
|
|
|
|
if (found)
|
|
|
|
{
|
|
|
|
if ((*found)->eq(item, 0))
|
|
|
|
continue; // Same field twice
|
|
|
|
if (report_error != IGNORE_ERRORS)
|
|
|
|
my_error(ER_NON_UNIQ_ERROR, MYF(0),
|
|
|
|
find->full_name(), current_thd->where);
|
|
|
|
return (Item**) 0;
|
|
|
|
}
|
|
|
|
found= li.ref();
|
|
|
|
*counter= i;
|
|
|
|
*resolution= fname_cmp ? RESOLVED_AGAINST_ALIAS:
|
|
|
|
RESOLVED_WITH_NO_ALIAS;
|
2004-08-20 15:53:46 +02:00
|
|
|
}
|
2007-03-05 04:54:35 +01:00
|
|
|
else if (!fname_cmp)
|
2004-08-20 15:53:46 +02:00
|
|
|
{
|
2007-03-05 04:54:35 +01:00
|
|
|
/*
|
|
|
|
We will use non-aliased field or react on such ambiguities only if
|
|
|
|
we won't be able to find aliased field.
|
|
|
|
Again if we have ambiguity with field outside of select list
|
|
|
|
we should prefer fields from select list.
|
|
|
|
*/
|
|
|
|
if (found_unaliased)
|
|
|
|
{
|
|
|
|
if ((*found_unaliased)->eq(item, 0))
|
|
|
|
continue; // Same field twice
|
|
|
|
found_unaliased_non_uniq= 1;
|
|
|
|
}
|
2004-08-20 15:53:46 +02:00
|
|
|
found_unaliased= li.ref();
|
|
|
|
unaliased_counter= i;
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
}
|
2007-03-05 04:54:35 +01:00
|
|
|
else if (!table_name)
|
|
|
|
{
|
|
|
|
if (is_ref_by_name && find->name && item->name &&
|
|
|
|
!my_strcasecmp(system_charset_info,item->name,find->name))
|
|
|
|
{
|
|
|
|
found= li.ref();
|
|
|
|
*counter= i;
|
|
|
|
*resolution= RESOLVED_AGAINST_ALIAS;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
else if (find->eq(item,0))
|
|
|
|
{
|
|
|
|
found= li.ref();
|
|
|
|
*counter= i;
|
|
|
|
*resolution= RESOLVED_IGNORING_ALIAS;
|
|
|
|
break;
|
|
|
|
}
|
2008-01-09 15:49:13 +01:00
|
|
|
}
|
|
|
|
else if (table_name && item->type() == Item::REF_ITEM &&
|
|
|
|
((Item_ref *)item)->ref_type() == Item_ref::VIEW_REF)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
TODO:Here we process prefixed view references only. What we should
|
|
|
|
really do is process all types of Item_refs. But this will currently
|
|
|
|
lead to a clash with the way references to outer SELECTs (from the
|
|
|
|
HAVING clause) are handled in e.g. :
|
|
|
|
SELECT 1 FROM t1 AS t1_o GROUP BY a
|
|
|
|
HAVING (SELECT t1_o.a FROM t1 AS t1_i GROUP BY t1_i.a LIMIT 1).
|
|
|
|
Processing all Item_refs here will cause t1_o.a to resolve to itself.
|
|
|
|
We still need to process the special case of Item_direct_view_ref
|
|
|
|
because in the context of views they have the same meaning as
|
|
|
|
Item_field for tables.
|
|
|
|
*/
|
|
|
|
Item_ident *item_ref= (Item_ident *) item;
|
|
|
|
if (item_ref->name && item_ref->table_name &&
|
|
|
|
!my_strcasecmp(system_charset_info, item_ref->name, field_name) &&
|
|
|
|
!my_strcasecmp(table_alias_charset, item_ref->table_name,
|
|
|
|
table_name) &&
|
|
|
|
(!db_name || (item_ref->db_name &&
|
|
|
|
!strcmp (item_ref->db_name, db_name))))
|
|
|
|
{
|
|
|
|
found= li.ref();
|
|
|
|
*counter= i;
|
|
|
|
*resolution= RESOLVED_IGNORING_ALIAS;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2004-08-20 15:53:46 +02:00
|
|
|
if (!found)
|
|
|
|
{
|
|
|
|
if (found_unaliased_non_uniq)
|
|
|
|
{
|
|
|
|
if (report_error != IGNORE_ERRORS)
|
2004-11-13 18:35:51 +01:00
|
|
|
my_error(ER_NON_UNIQ_ERROR, MYF(0),
|
|
|
|
find->full_name(), current_thd->where);
|
2004-08-20 15:53:46 +02:00
|
|
|
return (Item **) 0;
|
|
|
|
}
|
|
|
|
if (found_unaliased)
|
|
|
|
{
|
|
|
|
found= found_unaliased;
|
|
|
|
*counter= unaliased_counter;
|
2007-03-05 04:54:35 +01:00
|
|
|
*resolution= RESOLVED_BEHIND_ALIAS;
|
2004-08-20 15:53:46 +02:00
|
|
|
}
|
|
|
|
}
|
2002-10-08 13:50:12 +02:00
|
|
|
if (found)
|
|
|
|
return found;
|
2004-08-31 13:35:04 +02:00
|
|
|
if (report_error != REPORT_EXCEPT_NOT_FOUND)
|
2002-10-08 13:50:12 +02:00
|
|
|
{
|
|
|
|
if (report_error == REPORT_ALL_ERRORS)
|
2004-11-13 18:35:51 +01:00
|
|
|
my_error(ER_BAD_FIELD_ERROR, MYF(0),
|
|
|
|
find->full_name(), current_thd->where);
|
2002-10-08 13:50:12 +02:00
|
|
|
return (Item **) 0;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
return (Item **) not_found_item;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
Test if a string is a member of a list of strings.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
test_if_string_in_list()
|
|
|
|
find the string to look for
|
|
|
|
str_list a list of strings to be searched
|
|
|
|
|
|
|
|
DESCRIPTION
|
|
|
|
Sequentially search a list of strings for a string, and test whether
|
|
|
|
the list contains the same string.
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
TRUE if find is in str_list
|
|
|
|
FALSE otherwise
|
|
|
|
*/
|
|
|
|
|
|
|
|
static bool
|
|
|
|
test_if_string_in_list(const char *find, List<String> *str_list)
|
|
|
|
{
|
|
|
|
List_iterator<String> str_list_it(*str_list);
|
|
|
|
String *curr_str;
|
|
|
|
size_t find_length= strlen(find);
|
|
|
|
while ((curr_str= str_list_it++))
|
|
|
|
{
|
|
|
|
if (find_length != curr_str->length())
|
|
|
|
continue;
|
2005-09-12 07:28:53 +02:00
|
|
|
if (!my_strcasecmp(system_charset_info, find, curr_str->ptr()))
|
2005-08-12 16:57:19 +02:00
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Create a new name resolution context for an item so that it is
|
|
|
|
being resolved in a specific table reference.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
set_new_item_local_context()
|
|
|
|
thd pointer to current thread
|
|
|
|
item item for which new context is created and set
|
|
|
|
table_ref table ref where an item showld be resolved
|
|
|
|
|
|
|
|
DESCRIPTION
|
|
|
|
Create a new name resolution context for an item, so that the item
|
|
|
|
is resolved only the supplied 'table_ref'.
|
|
|
|
|
|
|
|
RETURN
|
2005-08-17 12:46:36 +02:00
|
|
|
FALSE if all OK
|
|
|
|
TRUE otherwise
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
|
|
|
|
|
|
|
static bool
|
|
|
|
set_new_item_local_context(THD *thd, Item_ident *item, TABLE_LIST *table_ref)
|
|
|
|
{
|
|
|
|
Name_resolution_context *context;
|
2005-08-22 00:13:37 +02:00
|
|
|
if (!(context= new (thd->mem_root) Name_resolution_context))
|
2005-08-12 16:57:19 +02:00
|
|
|
return TRUE;
|
|
|
|
context->init();
|
2005-08-17 12:46:36 +02:00
|
|
|
context->first_name_resolution_table=
|
|
|
|
context->last_name_resolution_table= table_ref;
|
2005-08-12 16:57:19 +02:00
|
|
|
item->context= context;
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Find and mark the common columns of two table references.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
mark_common_columns()
|
|
|
|
thd [in] current thread
|
|
|
|
table_ref_1 [in] the first (left) join operand
|
|
|
|
table_ref_2 [in] the second (right) join operand
|
|
|
|
using_fields [in] if the join is JOIN...USING - the join columns,
|
|
|
|
if NATURAL join, then NULL
|
|
|
|
found_using_fields [out] number of fields from the USING clause that were
|
|
|
|
found among the common fields
|
|
|
|
|
|
|
|
DESCRIPTION
|
|
|
|
The procedure finds the common columns of two relations (either
|
|
|
|
tables or intermediate join results), and adds an equi-join condition
|
|
|
|
to the ON clause of 'table_ref_2' for each pair of matching columns.
|
|
|
|
If some of table_ref_XXX represents a base table or view, then we
|
|
|
|
create new 'Natural_join_column' instances for each column
|
|
|
|
reference and store them in the 'join_columns' of the table
|
|
|
|
reference.
|
|
|
|
|
|
|
|
IMPLEMENTATION
|
|
|
|
The procedure assumes that store_natural_using_join_columns() was
|
|
|
|
called for the previous level of NATURAL/USING joins.
|
|
|
|
|
|
|
|
RETURN
|
2005-08-17 12:46:36 +02:00
|
|
|
TRUE error when some common column is non-unique, or out of memory
|
|
|
|
FALSE OK
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
|
|
|
|
|
|
|
static bool
|
|
|
|
mark_common_columns(THD *thd, TABLE_LIST *table_ref_1, TABLE_LIST *table_ref_2,
|
|
|
|
List<String> *using_fields, uint *found_using_fields)
|
|
|
|
{
|
|
|
|
Field_iterator_table_ref it_1, it_2;
|
|
|
|
Natural_join_column *nj_col_1, *nj_col_2;
|
|
|
|
Query_arena *arena, backup;
|
|
|
|
bool result= TRUE;
|
2006-03-02 10:50:15 +01:00
|
|
|
bool first_outer_loop= TRUE;
|
|
|
|
/*
|
|
|
|
Leaf table references to which new natural join columns are added
|
|
|
|
if the leaves are != NULL.
|
|
|
|
*/
|
|
|
|
TABLE_LIST *leaf_1= (table_ref_1->nested_join &&
|
|
|
|
!table_ref_1->is_natural_join) ?
|
|
|
|
NULL : table_ref_1;
|
|
|
|
TABLE_LIST *leaf_2= (table_ref_2->nested_join &&
|
|
|
|
!table_ref_2->is_natural_join) ?
|
|
|
|
NULL : table_ref_2;
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
DBUG_ENTER("mark_common_columns");
|
2005-08-17 12:46:36 +02:00
|
|
|
DBUG_PRINT("info", ("operand_1: %s operand_2: %s",
|
2005-08-12 16:57:19 +02:00
|
|
|
table_ref_1->alias, table_ref_2->alias));
|
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
*found_using_fields= 0;
|
2005-09-02 15:21:19 +02:00
|
|
|
arena= thd->activate_stmt_arena_if_needed(&backup);
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
for (it_1.set(table_ref_1); !it_1.end_of_fields(); it_1.next())
|
|
|
|
{
|
2005-08-17 12:46:36 +02:00
|
|
|
bool found= FALSE;
|
2006-02-23 14:21:43 +01:00
|
|
|
const char *field_name_1;
|
2007-01-31 15:04:38 +01:00
|
|
|
/* true if field_name_1 is a member of using_fields */
|
|
|
|
bool is_using_column_1;
|
2008-10-07 23:34:00 +02:00
|
|
|
if (!(nj_col_1= it_1.get_or_create_column_ref(thd, leaf_1)))
|
2005-08-12 16:57:19 +02:00
|
|
|
goto err;
|
|
|
|
field_name_1= nj_col_1->name();
|
2007-01-31 15:04:38 +01:00
|
|
|
is_using_column_1= using_fields &&
|
|
|
|
test_if_string_in_list(field_name_1, using_fields);
|
|
|
|
DBUG_PRINT ("info", ("field_name_1=%s.%s",
|
|
|
|
nj_col_1->table_name() ? nj_col_1->table_name() : "",
|
|
|
|
field_name_1));
|
2005-08-12 16:57:19 +02:00
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
/*
|
|
|
|
Find a field with the same name in table_ref_2.
|
|
|
|
|
|
|
|
Note that for the second loop, it_2.set() will iterate over
|
|
|
|
table_ref_2->join_columns and not generate any new elements or
|
|
|
|
lists.
|
|
|
|
*/
|
2005-08-12 16:57:19 +02:00
|
|
|
nj_col_2= NULL;
|
|
|
|
for (it_2.set(table_ref_2); !it_2.end_of_fields(); it_2.next())
|
|
|
|
{
|
|
|
|
Natural_join_column *cur_nj_col_2;
|
2005-08-17 12:46:36 +02:00
|
|
|
const char *cur_field_name_2;
|
2008-10-07 23:34:00 +02:00
|
|
|
if (!(cur_nj_col_2= it_2.get_or_create_column_ref(thd, leaf_2)))
|
2005-08-12 16:57:19 +02:00
|
|
|
goto err;
|
2005-08-17 12:46:36 +02:00
|
|
|
cur_field_name_2= cur_nj_col_2->name();
|
2007-01-31 15:04:38 +01:00
|
|
|
DBUG_PRINT ("info", ("cur_field_name_2=%s.%s",
|
|
|
|
cur_nj_col_2->table_name() ?
|
|
|
|
cur_nj_col_2->table_name() : "",
|
|
|
|
cur_field_name_2));
|
2005-08-12 16:57:19 +02:00
|
|
|
|
2005-09-08 10:29:52 +02:00
|
|
|
/*
|
|
|
|
Compare the two columns and check for duplicate common fields.
|
|
|
|
A common field is duplicate either if it was already found in
|
|
|
|
table_ref_2 (then found == TRUE), or if a field in table_ref_2
|
|
|
|
was already matched by some previous field in table_ref_1
|
|
|
|
(then cur_nj_col_2->is_common == TRUE).
|
2007-01-31 15:04:38 +01:00
|
|
|
Note that it is too early to check the columns outside of the
|
|
|
|
USING list for ambiguity because they are not actually "referenced"
|
|
|
|
here. These columns must be checked only on unqualified reference
|
|
|
|
by name (e.g. in SELECT list).
|
2005-09-08 10:29:52 +02:00
|
|
|
*/
|
2005-08-12 16:57:19 +02:00
|
|
|
if (!my_strcasecmp(system_charset_info, field_name_1, cur_field_name_2))
|
|
|
|
{
|
2007-01-31 15:04:38 +01:00
|
|
|
DBUG_PRINT ("info", ("match c1.is_common=%d", nj_col_1->is_common));
|
|
|
|
if (cur_nj_col_2->is_common ||
|
|
|
|
(found && (!using_fields || is_using_column_1)))
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
|
|
|
my_error(ER_NON_UNIQ_ERROR, MYF(0), field_name_1, thd->where);
|
|
|
|
goto err;
|
|
|
|
}
|
|
|
|
nj_col_2= cur_nj_col_2;
|
|
|
|
found= TRUE;
|
|
|
|
}
|
|
|
|
}
|
2006-03-02 10:50:15 +01:00
|
|
|
if (first_outer_loop && leaf_2)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
Make sure that the next inner loop "knows" that all columns
|
|
|
|
are materialized already.
|
|
|
|
*/
|
|
|
|
leaf_2->is_join_columns_complete= TRUE;
|
|
|
|
first_outer_loop= FALSE;
|
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
if (!found)
|
2005-08-17 12:46:36 +02:00
|
|
|
continue; // No matching field
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
field_1 and field_2 have the same names. Check if they are in the USING
|
|
|
|
clause (if present), mark them as common fields, and add a new
|
|
|
|
equi-join condition to the ON clause.
|
|
|
|
*/
|
2007-01-31 15:04:38 +01:00
|
|
|
if (nj_col_2 && (!using_fields ||is_using_column_1))
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
|
|
|
Item *item_1= nj_col_1->create_item(thd);
|
|
|
|
Item *item_2= nj_col_2->create_item(thd);
|
|
|
|
Field *field_1= nj_col_1->field();
|
|
|
|
Field *field_2= nj_col_2->field();
|
|
|
|
Item_ident *item_ident_1, *item_ident_2;
|
2005-08-17 12:46:36 +02:00
|
|
|
Item_func_eq *eq_cond;
|
|
|
|
|
|
|
|
if (!item_1 || !item_2)
|
|
|
|
goto err; // out of memory
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
/*
|
2005-08-17 12:46:36 +02:00
|
|
|
The following assert checks that the two created items are of
|
2005-08-12 16:57:19 +02:00
|
|
|
type Item_ident.
|
|
|
|
*/
|
|
|
|
DBUG_ASSERT(!thd->lex->current_select->no_wrap_view_item);
|
|
|
|
/*
|
|
|
|
In the case of no_wrap_view_item == 0, the created items must be
|
|
|
|
of sub-classes of Item_ident.
|
|
|
|
*/
|
|
|
|
DBUG_ASSERT(item_1->type() == Item::FIELD_ITEM ||
|
|
|
|
item_1->type() == Item::REF_ITEM);
|
|
|
|
DBUG_ASSERT(item_2->type() == Item::FIELD_ITEM ||
|
|
|
|
item_2->type() == Item::REF_ITEM);
|
|
|
|
|
|
|
|
/*
|
|
|
|
We need to cast item_1,2 to Item_ident, because we need to hook name
|
|
|
|
resolution contexts specific to each item.
|
|
|
|
*/
|
|
|
|
item_ident_1= (Item_ident*) item_1;
|
|
|
|
item_ident_2= (Item_ident*) item_2;
|
|
|
|
/*
|
|
|
|
Create and hook special name resolution contexts to each item in the
|
|
|
|
new join condition . We need this to both speed-up subsequent name
|
|
|
|
resolution of these items, and to enable proper name resolution of
|
|
|
|
the items during the execute phase of PS.
|
|
|
|
*/
|
2005-08-19 14:22:30 +02:00
|
|
|
if (set_new_item_local_context(thd, item_ident_1, nj_col_1->table_ref) ||
|
|
|
|
set_new_item_local_context(thd, item_ident_2, nj_col_2->table_ref))
|
2005-08-12 16:57:19 +02:00
|
|
|
goto err;
|
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
if (!(eq_cond= new Item_func_eq(item_ident_1, item_ident_2)))
|
|
|
|
goto err; /* Out of memory. */
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
Add the new equi-join condition to the ON clause. Notice that
|
|
|
|
fix_fields() is applied to all ON conditions in setup_conds()
|
|
|
|
so we don't do it here.
|
|
|
|
*/
|
2005-08-17 12:46:36 +02:00
|
|
|
add_join_on((table_ref_1->outer_join & JOIN_TYPE_RIGHT ?
|
|
|
|
table_ref_1 : table_ref_2),
|
|
|
|
eq_cond);
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
nj_col_1->is_common= nj_col_2->is_common= TRUE;
|
2007-01-31 15:04:38 +01:00
|
|
|
DBUG_PRINT ("info", ("%s.%s and %s.%s are common",
|
|
|
|
nj_col_1->table_name() ?
|
|
|
|
nj_col_1->table_name() : "",
|
|
|
|
nj_col_1->name(),
|
|
|
|
nj_col_2->table_name() ?
|
|
|
|
nj_col_2->table_name() : "",
|
|
|
|
nj_col_2->name()));
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
if (field_1)
|
|
|
|
{
|
2005-08-25 21:02:40 +02:00
|
|
|
TABLE *table_1= nj_col_1->table_ref->table;
|
2005-08-12 16:57:19 +02:00
|
|
|
/* Mark field_1 used for table cache. */
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
bitmap_set_bit(table_1->read_set, field_1->field_index);
|
2007-03-05 18:08:41 +01:00
|
|
|
table_1->covering_keys.intersect(field_1->part_of_key);
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
table_1->merge_keys.merge(field_1->part_of_key);
|
2005-08-12 16:57:19 +02:00
|
|
|
}
|
|
|
|
if (field_2)
|
|
|
|
{
|
2005-08-25 21:02:40 +02:00
|
|
|
TABLE *table_2= nj_col_2->table_ref->table;
|
2005-08-12 16:57:19 +02:00
|
|
|
/* Mark field_2 used for table cache. */
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
bitmap_set_bit(table_2->read_set, field_2->field_index);
|
2007-03-05 18:08:41 +01:00
|
|
|
table_2->covering_keys.intersect(field_2->part_of_key);
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
table_2->merge_keys.merge(field_2->part_of_key);
|
2005-08-12 16:57:19 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (using_fields != NULL)
|
|
|
|
++(*found_using_fields);
|
|
|
|
}
|
|
|
|
}
|
2006-03-02 10:50:15 +01:00
|
|
|
if (leaf_1)
|
|
|
|
leaf_1->is_join_columns_complete= TRUE;
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
Everything is OK.
|
|
|
|
Notice that at this point there may be some column names in the USING
|
|
|
|
clause that are not among the common columns. This is an SQL error and
|
|
|
|
we check for this error in store_natural_using_join_columns() when
|
|
|
|
(found_using_fields < length(join_using_fields)).
|
|
|
|
*/
|
|
|
|
result= FALSE;
|
|
|
|
|
|
|
|
err:
|
|
|
|
if (arena)
|
2005-09-02 15:21:19 +02:00
|
|
|
thd->restore_active_arena(arena, &backup);
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN(result);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Materialize and store the row type of NATURAL/USING join.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
store_natural_using_join_columns()
|
|
|
|
thd current thread
|
|
|
|
natural_using_join the table reference of the NATURAL/USING join
|
|
|
|
table_ref_1 the first (left) operand (of a NATURAL/USING join).
|
|
|
|
table_ref_2 the second (right) operand (of a NATURAL/USING join).
|
|
|
|
using_fields if the join is JOIN...USING - the join columns,
|
|
|
|
if NATURAL join, then NULL
|
|
|
|
found_using_fields number of fields from the USING clause that were
|
|
|
|
found among the common fields
|
|
|
|
|
|
|
|
DESCRIPTION
|
|
|
|
Iterate over the columns of both join operands and sort and store
|
|
|
|
all columns into the 'join_columns' list of natural_using_join
|
|
|
|
where the list is formed by three parts:
|
|
|
|
part1: The coalesced columns of table_ref_1 and table_ref_2,
|
|
|
|
sorted according to the column order of the first table.
|
|
|
|
part2: The other columns of the first table, in the order in
|
|
|
|
which they were defined in CREATE TABLE.
|
|
|
|
part3: The other columns of the second table, in the order in
|
|
|
|
which they were defined in CREATE TABLE.
|
|
|
|
Time complexity - O(N1+N2), where Ni = length(table_ref_i).
|
|
|
|
|
|
|
|
IMPLEMENTATION
|
|
|
|
The procedure assumes that mark_common_columns() has been called
|
|
|
|
for the join that is being processed.
|
|
|
|
|
|
|
|
RETURN
|
2005-08-17 12:46:36 +02:00
|
|
|
TRUE error: Some common column is ambiguous
|
|
|
|
FALSE OK
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
|
|
|
|
|
|
|
static bool
|
|
|
|
store_natural_using_join_columns(THD *thd, TABLE_LIST *natural_using_join,
|
|
|
|
TABLE_LIST *table_ref_1,
|
|
|
|
TABLE_LIST *table_ref_2,
|
|
|
|
List<String> *using_fields,
|
|
|
|
uint found_using_fields)
|
|
|
|
{
|
|
|
|
Field_iterator_table_ref it_1, it_2;
|
|
|
|
Natural_join_column *nj_col_1, *nj_col_2;
|
|
|
|
Query_arena *arena, backup;
|
|
|
|
bool result= TRUE;
|
2005-08-17 12:46:36 +02:00
|
|
|
List<Natural_join_column> *non_join_columns;
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_ENTER("store_natural_using_join_columns");
|
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
DBUG_ASSERT(!natural_using_join->join_columns);
|
|
|
|
|
2005-09-02 15:21:19 +02:00
|
|
|
arena= thd->activate_stmt_arena_if_needed(&backup);
|
2005-08-12 16:57:19 +02:00
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
if (!(non_join_columns= new List<Natural_join_column>) ||
|
|
|
|
!(natural_using_join->join_columns= new List<Natural_join_column>))
|
2005-08-12 16:57:19 +02:00
|
|
|
goto err;
|
|
|
|
|
|
|
|
/* Append the columns of the first join operand. */
|
|
|
|
for (it_1.set(table_ref_1); !it_1.end_of_fields(); it_1.next())
|
|
|
|
{
|
2005-11-28 20:57:50 +01:00
|
|
|
nj_col_1= it_1.get_natural_column_ref();
|
2005-08-12 16:57:19 +02:00
|
|
|
if (nj_col_1->is_common)
|
|
|
|
{
|
|
|
|
natural_using_join->join_columns->push_back(nj_col_1);
|
|
|
|
/* Reset the common columns for the next call to mark_common_columns. */
|
|
|
|
nj_col_1->is_common= FALSE;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
non_join_columns->push_back(nj_col_1);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
Check that all columns in the USING clause are among the common
|
|
|
|
columns. If this is not the case, report the first one that was
|
|
|
|
not found in an error.
|
|
|
|
*/
|
|
|
|
if (using_fields && found_using_fields < using_fields->elements)
|
|
|
|
{
|
|
|
|
String *using_field_name;
|
|
|
|
List_iterator_fast<String> using_fields_it(*using_fields);
|
|
|
|
while ((using_field_name= using_fields_it++))
|
|
|
|
{
|
2005-08-17 12:46:36 +02:00
|
|
|
const char *using_field_name_ptr= using_field_name->c_ptr();
|
2005-08-12 16:57:19 +02:00
|
|
|
List_iterator_fast<Natural_join_column>
|
|
|
|
it(*(natural_using_join->join_columns));
|
|
|
|
Natural_join_column *common_field;
|
2005-08-17 12:46:36 +02:00
|
|
|
|
|
|
|
for (;;)
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
2005-08-17 12:46:36 +02:00
|
|
|
/* If reached the end of fields, and none was found, report error. */
|
|
|
|
if (!(common_field= it++))
|
|
|
|
{
|
|
|
|
my_error(ER_BAD_FIELD_ERROR, MYF(0), using_field_name_ptr,
|
|
|
|
current_thd->where);
|
|
|
|
goto err;
|
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
if (!my_strcasecmp(system_charset_info,
|
|
|
|
common_field->name(), using_field_name_ptr))
|
2005-08-17 12:46:36 +02:00
|
|
|
break; // Found match
|
2005-08-12 16:57:19 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Append the non-equi-join columns of the second join operand. */
|
|
|
|
for (it_2.set(table_ref_2); !it_2.end_of_fields(); it_2.next())
|
|
|
|
{
|
2005-11-28 20:57:50 +01:00
|
|
|
nj_col_2= it_2.get_natural_column_ref();
|
2005-08-12 16:57:19 +02:00
|
|
|
if (!nj_col_2->is_common)
|
|
|
|
non_join_columns->push_back(nj_col_2);
|
|
|
|
else
|
2005-08-17 12:46:36 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
/* Reset the common columns for the next call to mark_common_columns. */
|
|
|
|
nj_col_2->is_common= FALSE;
|
2005-08-17 12:46:36 +02:00
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (non_join_columns->elements > 0)
|
|
|
|
natural_using_join->join_columns->concat(non_join_columns);
|
|
|
|
natural_using_join->is_join_columns_complete= TRUE;
|
|
|
|
|
|
|
|
result= FALSE;
|
|
|
|
|
|
|
|
err:
|
|
|
|
if (arena)
|
2005-09-02 15:21:19 +02:00
|
|
|
thd->restore_active_arena(arena, &backup);
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN(result);
|
|
|
|
}
|
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
/*
|
|
|
|
Precompute and store the row types of the top-most NATURAL/USING joins.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
store_top_level_join_columns()
|
|
|
|
thd current thread
|
|
|
|
table_ref nested join or table in a FROM clause
|
|
|
|
left_neighbor neighbor table reference to the left of table_ref at the
|
|
|
|
same level in the join tree
|
|
|
|
right_neighbor neighbor table reference to the right of table_ref at the
|
|
|
|
same level in the join tree
|
|
|
|
|
|
|
|
DESCRIPTION
|
|
|
|
The procedure performs a post-order traversal of a nested join tree
|
|
|
|
and materializes the row types of NATURAL/USING joins in a
|
|
|
|
bottom-up manner until it reaches the TABLE_LIST elements that
|
|
|
|
represent the top-most NATURAL/USING joins. The procedure should be
|
|
|
|
applied to each element of SELECT_LEX::top_join_list (i.e. to each
|
|
|
|
top-level element of the FROM clause).
|
|
|
|
|
|
|
|
IMPLEMENTATION
|
|
|
|
Notice that the table references in the list nested_join->join_list
|
|
|
|
are in reverse order, thus when we iterate over it, we are moving
|
|
|
|
from the right to the left in the FROM clause.
|
|
|
|
|
|
|
|
RETURN
|
2005-08-17 12:46:36 +02:00
|
|
|
TRUE Error
|
|
|
|
FALSE OK
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
|
|
|
|
|
|
|
static bool
|
|
|
|
store_top_level_join_columns(THD *thd, TABLE_LIST *table_ref,
|
|
|
|
TABLE_LIST *left_neighbor,
|
|
|
|
TABLE_LIST *right_neighbor)
|
|
|
|
{
|
2005-08-24 09:51:28 +02:00
|
|
|
Query_arena *arena, backup;
|
|
|
|
bool result= TRUE;
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_ENTER("store_top_level_join_columns");
|
2005-08-17 12:46:36 +02:00
|
|
|
|
2005-09-02 15:21:19 +02:00
|
|
|
arena= thd->activate_stmt_arena_if_needed(&backup);
|
2005-08-24 09:51:28 +02:00
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
/* Call the procedure recursively for each nested table reference. */
|
|
|
|
if (table_ref->nested_join)
|
|
|
|
{
|
|
|
|
List_iterator_fast<TABLE_LIST> nested_it(table_ref->nested_join->join_list);
|
2006-07-21 10:59:46 +02:00
|
|
|
TABLE_LIST *same_level_left_neighbor= nested_it++;
|
|
|
|
TABLE_LIST *same_level_right_neighbor= NULL;
|
|
|
|
/* Left/right-most neighbors, possibly at higher levels in the join tree. */
|
|
|
|
TABLE_LIST *real_left_neighbor, *real_right_neighbor;
|
2005-08-17 12:46:36 +02:00
|
|
|
|
2006-07-21 10:59:46 +02:00
|
|
|
while (same_level_left_neighbor)
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
2006-07-21 10:59:46 +02:00
|
|
|
TABLE_LIST *cur_table_ref= same_level_left_neighbor;
|
|
|
|
same_level_left_neighbor= nested_it++;
|
2005-08-17 16:19:31 +02:00
|
|
|
/*
|
|
|
|
The order of RIGHT JOIN operands is reversed in 'join list' to
|
|
|
|
transform it into a LEFT JOIN. However, in this procedure we need
|
|
|
|
the join operands in their lexical order, so below we reverse the
|
2006-07-21 10:59:46 +02:00
|
|
|
join operands. Notice that this happens only in the first loop,
|
|
|
|
and not in the second one, as in the second loop
|
|
|
|
same_level_left_neighbor == NULL.
|
|
|
|
This is the correct behavior, because the second loop sets
|
|
|
|
cur_table_ref reference correctly after the join operands are
|
2005-08-17 16:19:31 +02:00
|
|
|
swapped in the first loop.
|
|
|
|
*/
|
2006-07-21 10:59:46 +02:00
|
|
|
if (same_level_left_neighbor &&
|
2005-08-17 16:19:31 +02:00
|
|
|
cur_table_ref->outer_join & JOIN_TYPE_RIGHT)
|
|
|
|
{
|
|
|
|
/* This can happen only for JOIN ... ON. */
|
|
|
|
DBUG_ASSERT(table_ref->nested_join->join_list.elements == 2);
|
2006-07-21 10:59:46 +02:00
|
|
|
swap_variables(TABLE_LIST*, same_level_left_neighbor, cur_table_ref);
|
2005-08-17 16:19:31 +02:00
|
|
|
}
|
|
|
|
|
2006-07-21 10:59:46 +02:00
|
|
|
/*
|
|
|
|
Pick the parent's left and right neighbors if there are no immediate
|
|
|
|
neighbors at the same level.
|
|
|
|
*/
|
|
|
|
real_left_neighbor= (same_level_left_neighbor) ?
|
|
|
|
same_level_left_neighbor : left_neighbor;
|
|
|
|
real_right_neighbor= (same_level_right_neighbor) ?
|
|
|
|
same_level_right_neighbor : right_neighbor;
|
|
|
|
|
2005-08-17 16:19:31 +02:00
|
|
|
if (cur_table_ref->nested_join &&
|
|
|
|
store_top_level_join_columns(thd, cur_table_ref,
|
2006-07-21 10:59:46 +02:00
|
|
|
real_left_neighbor, real_right_neighbor))
|
2005-08-24 09:51:28 +02:00
|
|
|
goto err;
|
2006-07-21 10:59:46 +02:00
|
|
|
same_level_right_neighbor= cur_table_ref;
|
2005-08-12 16:57:19 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
If this is a NATURAL/USING join, materialize its result columns and
|
|
|
|
convert to a JOIN ... ON.
|
|
|
|
*/
|
|
|
|
if (table_ref->is_natural_join)
|
|
|
|
{
|
|
|
|
DBUG_ASSERT(table_ref->nested_join &&
|
|
|
|
table_ref->nested_join->join_list.elements == 2);
|
|
|
|
List_iterator_fast<TABLE_LIST> operand_it(table_ref->nested_join->join_list);
|
|
|
|
/*
|
|
|
|
Notice that the order of join operands depends on whether table_ref
|
|
|
|
represents a LEFT or a RIGHT join. In a RIGHT join, the operands are
|
|
|
|
in inverted order.
|
|
|
|
*/
|
|
|
|
TABLE_LIST *table_ref_2= operand_it++; /* Second NATURAL join operand.*/
|
|
|
|
TABLE_LIST *table_ref_1= operand_it++; /* First NATURAL join operand. */
|
|
|
|
List<String> *using_fields= table_ref->join_using_fields;
|
|
|
|
uint found_using_fields;
|
|
|
|
|
|
|
|
/*
|
|
|
|
The two join operands were interchanged in the parser, change the order
|
|
|
|
back for 'mark_common_columns'.
|
|
|
|
*/
|
|
|
|
if (table_ref_2->outer_join & JOIN_TYPE_RIGHT)
|
|
|
|
swap_variables(TABLE_LIST*, table_ref_1, table_ref_2);
|
|
|
|
if (mark_common_columns(thd, table_ref_1, table_ref_2,
|
|
|
|
using_fields, &found_using_fields))
|
2005-08-24 09:51:28 +02:00
|
|
|
goto err;
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
Swap the join operands back, so that we pick the columns of the second
|
|
|
|
one as the coalesced columns. In this way the coalesced columns are the
|
|
|
|
same as of an equivalent LEFT JOIN.
|
|
|
|
*/
|
|
|
|
if (table_ref_1->outer_join & JOIN_TYPE_RIGHT)
|
|
|
|
swap_variables(TABLE_LIST*, table_ref_1, table_ref_2);
|
|
|
|
if (store_natural_using_join_columns(thd, table_ref, table_ref_1,
|
|
|
|
table_ref_2, using_fields,
|
|
|
|
found_using_fields))
|
2005-08-24 09:51:28 +02:00
|
|
|
goto err;
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
Change NATURAL JOIN to JOIN ... ON. We do this for both operands
|
|
|
|
because either one of them or the other is the one with the
|
|
|
|
natural join flag because RIGHT joins are transformed into LEFT,
|
|
|
|
and the two tables may be reordered.
|
|
|
|
*/
|
|
|
|
table_ref_1->natural_join= table_ref_2->natural_join= NULL;
|
|
|
|
|
2005-08-23 19:03:32 +02:00
|
|
|
/* Add a TRUE condition to outer joins that have no common columns. */
|
|
|
|
if (table_ref_2->outer_join &&
|
|
|
|
!table_ref_1->on_expr && !table_ref_2->on_expr)
|
|
|
|
table_ref_2->on_expr= new Item_int((longlong) 1,1); /* Always true. */
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
/* Change this table reference to become a leaf for name resolution. */
|
|
|
|
if (left_neighbor)
|
|
|
|
{
|
2005-08-17 12:46:36 +02:00
|
|
|
TABLE_LIST *last_leaf_on_the_left;
|
2005-08-12 16:57:19 +02:00
|
|
|
last_leaf_on_the_left= left_neighbor->last_leaf_for_name_resolution();
|
|
|
|
last_leaf_on_the_left->next_name_resolution_table= table_ref;
|
|
|
|
}
|
|
|
|
if (right_neighbor)
|
|
|
|
{
|
2005-08-17 12:46:36 +02:00
|
|
|
TABLE_LIST *first_leaf_on_the_right;
|
2005-08-12 16:57:19 +02:00
|
|
|
first_leaf_on_the_right= right_neighbor->first_leaf_for_name_resolution();
|
|
|
|
table_ref->next_name_resolution_table= first_leaf_on_the_right;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
table_ref->next_name_resolution_table= NULL;
|
|
|
|
}
|
2005-08-24 09:51:28 +02:00
|
|
|
result= FALSE; /* All is OK. */
|
|
|
|
|
|
|
|
err:
|
|
|
|
if (arena)
|
2005-09-02 15:21:19 +02:00
|
|
|
thd->restore_active_arena(arena, &backup);
|
2005-08-24 09:51:28 +02:00
|
|
|
DBUG_RETURN(result);
|
2005-08-12 16:57:19 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Compute and store the row types of the top-most NATURAL/USING joins
|
|
|
|
in a FROM clause.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
setup_natural_join_row_types()
|
|
|
|
thd current thread
|
|
|
|
from_clause list of top-level table references in a FROM clause
|
|
|
|
|
|
|
|
DESCRIPTION
|
|
|
|
Apply the procedure 'store_top_level_join_columns' to each of the
|
|
|
|
top-level table referencs of the FROM clause. Adjust the list of tables
|
|
|
|
for name resolution - context->first_name_resolution_table to the
|
|
|
|
top-most, lef-most NATURAL/USING join.
|
|
|
|
|
|
|
|
IMPLEMENTATION
|
|
|
|
Notice that the table references in 'from_clause' are in reverse
|
|
|
|
order, thus when we iterate over it, we are moving from the right
|
|
|
|
to the left in the FROM clause.
|
|
|
|
|
|
|
|
RETURN
|
2005-08-17 12:46:36 +02:00
|
|
|
TRUE Error
|
|
|
|
FALSE OK
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
2005-08-17 12:46:36 +02:00
|
|
|
static bool setup_natural_join_row_types(THD *thd,
|
|
|
|
List<TABLE_LIST> *from_clause,
|
2005-08-12 16:57:19 +02:00
|
|
|
Name_resolution_context *context)
|
|
|
|
{
|
|
|
|
thd->where= "from clause";
|
|
|
|
if (from_clause->elements == 0)
|
|
|
|
return FALSE; /* We come here in the case of UNIONs. */
|
|
|
|
|
|
|
|
List_iterator_fast<TABLE_LIST> table_ref_it(*from_clause);
|
|
|
|
TABLE_LIST *table_ref; /* Current table reference. */
|
|
|
|
/* Table reference to the left of the current. */
|
2005-08-17 12:46:36 +02:00
|
|
|
TABLE_LIST *left_neighbor;
|
2005-08-12 16:57:19 +02:00
|
|
|
/* Table reference to the right of the current. */
|
|
|
|
TABLE_LIST *right_neighbor= NULL;
|
2008-10-07 23:34:00 +02:00
|
|
|
bool save_first_natural_join_processing=
|
|
|
|
context->select_lex->first_natural_join_processing;
|
|
|
|
|
|
|
|
context->select_lex->first_natural_join_processing= FALSE;
|
2005-08-12 16:57:19 +02:00
|
|
|
|
2005-08-17 12:46:36 +02:00
|
|
|
/* Note that tables in the list are in reversed order */
|
|
|
|
for (left_neighbor= table_ref_it++; left_neighbor ; )
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
|
|
|
table_ref= left_neighbor;
|
|
|
|
left_neighbor= table_ref_it++;
|
2008-10-07 23:34:00 +02:00
|
|
|
/*
|
|
|
|
Do not redo work if already done:
|
|
|
|
1) for stored procedures,
|
|
|
|
2) for multitable update after lock failure and table reopening.
|
|
|
|
*/
|
|
|
|
if (save_first_natural_join_processing)
|
2005-08-12 16:57:19 +02:00
|
|
|
{
|
2008-10-07 23:34:00 +02:00
|
|
|
context->select_lex->first_natural_join_processing= FALSE;
|
2006-06-08 12:34:03 +02:00
|
|
|
if (store_top_level_join_columns(thd, table_ref,
|
|
|
|
left_neighbor, right_neighbor))
|
|
|
|
return TRUE;
|
|
|
|
if (left_neighbor)
|
|
|
|
{
|
|
|
|
TABLE_LIST *first_leaf_on_the_right;
|
|
|
|
first_leaf_on_the_right= table_ref->first_leaf_for_name_resolution();
|
|
|
|
left_neighbor->next_name_resolution_table= first_leaf_on_the_right;
|
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
}
|
|
|
|
right_neighbor= table_ref;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
Store the top-most, left-most NATURAL/USING join, so that we start
|
|
|
|
the search from that one instead of context->table_list. At this point
|
2005-08-17 12:46:36 +02:00
|
|
|
right_neighbor points to the left-most top-level table reference in the
|
2005-08-12 16:57:19 +02:00
|
|
|
FROM clause.
|
|
|
|
*/
|
|
|
|
DBUG_ASSERT(right_neighbor);
|
|
|
|
context->first_name_resolution_table=
|
|
|
|
right_neighbor->first_leaf_for_name_resolution();
|
|
|
|
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
/****************************************************************************
|
2003-01-25 01:25:52 +01:00
|
|
|
** Expand all '*' in given fields
|
2000-07-31 21:29:14 +02:00
|
|
|
****************************************************************************/
|
|
|
|
|
2003-01-25 01:25:52 +01:00
|
|
|
int setup_wild(THD *thd, TABLE_LIST *tables, List<Item> &fields,
|
|
|
|
List<Item> *sum_func_list,
|
|
|
|
uint wild_num)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2003-01-25 01:25:52 +01:00
|
|
|
if (!wild_num)
|
2004-11-08 16:53:32 +01:00
|
|
|
return(0);
|
2004-10-26 18:30:01 +02:00
|
|
|
|
2004-11-08 16:53:32 +01:00
|
|
|
Item *item;
|
2004-11-08 00:13:54 +01:00
|
|
|
List_iterator<Item> it(fields);
|
2005-06-15 19:58:35 +02:00
|
|
|
Query_arena *arena, backup;
|
2004-11-08 16:53:32 +01:00
|
|
|
DBUG_ENTER("setup_wild");
|
2004-06-25 15:52:01 +02:00
|
|
|
|
|
|
|
/*
|
2004-09-09 05:59:26 +02:00
|
|
|
Don't use arena if we are not in prepared statements or stored procedures
|
|
|
|
For PS/SP we have to use arena to remember the changes
|
2004-06-25 15:52:01 +02:00
|
|
|
*/
|
2005-09-02 15:21:19 +02:00
|
|
|
arena= thd->activate_stmt_arena_if_needed(&backup);
|
2004-08-21 00:02:46 +02:00
|
|
|
|
2007-04-24 20:35:57 +02:00
|
|
|
thd->lex->current_select->cur_pos_in_select_list= 0;
|
2004-09-06 14:14:10 +02:00
|
|
|
while (wild_num && (item= it++))
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2004-10-26 18:30:01 +02:00
|
|
|
if (item->type() == Item::FIELD_ITEM &&
|
|
|
|
((Item_field*) item)->field_name &&
|
2003-12-11 05:24:08 +01:00
|
|
|
((Item_field*) item)->field_name[0] == '*' &&
|
|
|
|
!((Item_field*) item)->field)
|
2001-03-21 19:13:46 +01:00
|
|
|
{
|
2002-11-25 00:38:56 +01:00
|
|
|
uint elem= fields.elements;
|
2004-07-16 00:15:55 +02:00
|
|
|
bool any_privileges= ((Item_field *) item)->any_privileges;
|
2004-06-10 20:33:15 +02:00
|
|
|
Item_subselect *subsel= thd->lex->current_select->master_unit()->item;
|
|
|
|
if (subsel &&
|
|
|
|
subsel->substype() == Item_subselect::EXISTS_SUBS)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
It is EXISTS(SELECT * ...) and we can replace * by any constant.
|
|
|
|
|
|
|
|
Item_int do not need fix_fields() because it is basic constant.
|
|
|
|
*/
|
2007-03-09 06:05:08 +01:00
|
|
|
it.replace(new Item_int("Not_used", (longlong) 1,
|
|
|
|
MY_INT64_NUM_DECIMAL_DIGITS));
|
2004-06-10 20:33:15 +02:00
|
|
|
}
|
2005-07-01 06:05:42 +02:00
|
|
|
else if (insert_fields(thd, ((Item_field*) item)->context,
|
|
|
|
((Item_field*) item)->db_name,
|
2004-09-10 01:22:44 +02:00
|
|
|
((Item_field*) item)->table_name, &it,
|
2005-06-21 19:15:21 +02:00
|
|
|
any_privileges))
|
2004-06-25 15:52:01 +02:00
|
|
|
{
|
|
|
|
if (arena)
|
2005-09-02 15:21:19 +02:00
|
|
|
thd->restore_active_arena(arena, &backup);
|
2004-10-26 18:30:01 +02:00
|
|
|
DBUG_RETURN(-1);
|
2004-06-25 15:52:01 +02:00
|
|
|
}
|
2002-09-03 22:00:23 +02:00
|
|
|
if (sum_func_list)
|
2002-09-05 15:17:08 +02:00
|
|
|
{
|
|
|
|
/*
|
|
|
|
sum_func_list is a list that has the fields list as a tail.
|
|
|
|
Because of this we have to update the element count also for this
|
|
|
|
list after expanding the '*' entry.
|
|
|
|
*/
|
|
|
|
sum_func_list->elements+= fields.elements - elem;
|
|
|
|
}
|
2003-01-25 01:25:52 +01:00
|
|
|
wild_num--;
|
2001-03-21 19:13:46 +01:00
|
|
|
}
|
2007-04-24 20:35:57 +02:00
|
|
|
else
|
|
|
|
thd->lex->current_select->cur_pos_in_select_list++;
|
2001-03-21 19:13:46 +01:00
|
|
|
}
|
2007-04-24 20:35:57 +02:00
|
|
|
thd->lex->current_select->cur_pos_in_select_list= UNDEF_POS;
|
2004-06-25 15:52:01 +02:00
|
|
|
if (arena)
|
|
|
|
{
|
|
|
|
/* make * substituting permanent */
|
|
|
|
SELECT_LEX *select_lex= thd->lex->current_select;
|
|
|
|
select_lex->with_wild= 0;
|
2009-07-16 14:19:22 +02:00
|
|
|
/*
|
|
|
|
The assignment below is translated to memcpy() call (at least on some
|
|
|
|
platforms). memcpy() expects that source and destination areas do not
|
|
|
|
overlap. That problem was detected by valgrind.
|
|
|
|
*/
|
|
|
|
if (&select_lex->item_list != &fields)
|
|
|
|
select_lex->item_list= fields;
|
2004-06-25 15:52:01 +02:00
|
|
|
|
2005-09-02 15:21:19 +02:00
|
|
|
thd->restore_active_arena(arena, &backup);
|
2004-06-25 15:52:01 +02:00
|
|
|
}
|
2004-10-26 18:30:01 +02:00
|
|
|
DBUG_RETURN(0);
|
2001-03-21 19:13:46 +01:00
|
|
|
}
|
|
|
|
|
2003-01-25 01:25:52 +01:00
|
|
|
/****************************************************************************
|
|
|
|
** Check that all given fields exists and fill struct with current data
|
|
|
|
****************************************************************************/
|
|
|
|
|
2005-07-01 06:05:42 +02:00
|
|
|
bool setup_fields(THD *thd, Item **ref_pointer_array,
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
List<Item> &fields, enum_mark_columns mark_used_columns,
|
2004-10-20 03:04:37 +02:00
|
|
|
List<Item> *sum_func_list, bool allow_sum_func)
|
2003-01-25 01:25:52 +01:00
|
|
|
{
|
|
|
|
reg2 Item *item;
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
enum_mark_columns save_mark_used_columns= thd->mark_used_columns;
|
2005-10-15 23:32:37 +02:00
|
|
|
nesting_map save_allow_sum_func= thd->lex->allow_sum_func;
|
2003-01-25 01:25:52 +01:00
|
|
|
List_iterator<Item> it(fields);
|
2007-03-03 22:47:42 +01:00
|
|
|
bool save_is_item_list_lookup;
|
2003-01-25 01:25:52 +01:00
|
|
|
DBUG_ENTER("setup_fields");
|
|
|
|
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
thd->mark_used_columns= mark_used_columns;
|
|
|
|
DBUG_PRINT("info", ("thd->mark_used_columns: %d", thd->mark_used_columns));
|
2005-10-15 23:32:37 +02:00
|
|
|
if (allow_sum_func)
|
|
|
|
thd->lex->allow_sum_func|= 1 << thd->lex->current_select->nest_level;
|
2005-10-25 11:02:48 +02:00
|
|
|
thd->where= THD::DEFAULT_WHERE;
|
2007-03-03 22:47:42 +01:00
|
|
|
save_is_item_list_lookup= thd->lex->current_select->is_item_list_lookup;
|
|
|
|
thd->lex->current_select->is_item_list_lookup= 0;
|
2003-01-25 01:25:52 +01:00
|
|
|
|
2005-01-24 16:17:19 +01:00
|
|
|
/*
|
|
|
|
To prevent fail on forward lookup we fill it with zerows,
|
|
|
|
then if we got pointer on zero after find_item_in_list we will know
|
|
|
|
that it is forward lookup.
|
|
|
|
|
|
|
|
There is other way to solve problem: fill array with pointers to list,
|
|
|
|
but it will be slower.
|
|
|
|
|
|
|
|
TODO: remove it when (if) we made one list for allfields and
|
|
|
|
ref_pointer_array
|
|
|
|
*/
|
|
|
|
if (ref_pointer_array)
|
|
|
|
bzero(ref_pointer_array, sizeof(Item *) * fields.elements);
|
|
|
|
|
2008-09-18 10:38:44 +02:00
|
|
|
/*
|
|
|
|
We call set_entry() there (before fix_fields() of the whole list of field
|
|
|
|
items) because:
|
|
|
|
1) the list of field items has same order as in the query, and the
|
|
|
|
Item_func_get_user_var item may go before the Item_func_set_user_var:
|
|
|
|
SELECT @a, @a := 10 FROM t;
|
|
|
|
2) The entry->update_query_id value controls constantness of
|
|
|
|
Item_func_get_user_var items, so in presence of Item_func_set_user_var
|
|
|
|
items we have to refresh their entries before fixing of
|
|
|
|
Item_func_get_user_var items.
|
|
|
|
*/
|
|
|
|
List_iterator<Item_func_set_user_var> li(thd->lex->set_var_list);
|
|
|
|
Item_func_set_user_var *var;
|
|
|
|
while ((var= li++))
|
|
|
|
var->set_entry(thd, FALSE);
|
|
|
|
|
2003-02-02 22:30:01 +01:00
|
|
|
Item **ref= ref_pointer_array;
|
2007-01-11 21:18:01 +01:00
|
|
|
thd->lex->current_select->cur_pos_in_select_list= 0;
|
2003-02-02 22:30:01 +01:00
|
|
|
while ((item= it++))
|
2003-01-25 01:25:52 +01:00
|
|
|
{
|
2009-06-10 16:04:07 +02:00
|
|
|
if ((!item->fixed && item->fix_fields(thd, it.ref())) ||
|
2004-06-25 15:52:01 +02:00
|
|
|
(item= *(it.ref()))->check_cols(1))
|
2004-07-23 08:20:58 +02:00
|
|
|
{
|
2007-03-03 22:47:42 +01:00
|
|
|
thd->lex->current_select->is_item_list_lookup= save_is_item_list_lookup;
|
2005-10-15 23:32:37 +02:00
|
|
|
thd->lex->allow_sum_func= save_allow_sum_func;
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
thd->mark_used_columns= save_mark_used_columns;
|
|
|
|
DBUG_PRINT("info", ("thd->mark_used_columns: %d", thd->mark_used_columns));
|
2004-10-20 03:04:37 +02:00
|
|
|
DBUG_RETURN(TRUE); /* purecov: inspected */
|
2004-07-23 08:20:58 +02:00
|
|
|
}
|
2003-02-02 22:30:01 +01:00
|
|
|
if (ref)
|
|
|
|
*(ref++)= item;
|
2003-01-25 01:25:52 +01:00
|
|
|
if (item->with_sum_func && item->type() != Item::SUM_FUNC_ITEM &&
|
|
|
|
sum_func_list)
|
2004-10-08 17:13:09 +02:00
|
|
|
item->split_sum_func(thd, ref_pointer_array, *sum_func_list);
|
2005-08-12 01:10:34 +02:00
|
|
|
thd->used_tables|= item->used_tables();
|
2007-01-11 21:18:01 +01:00
|
|
|
thd->lex->current_select->cur_pos_in_select_list++;
|
2003-01-25 01:25:52 +01:00
|
|
|
}
|
2007-03-03 22:47:42 +01:00
|
|
|
thd->lex->current_select->is_item_list_lookup= save_is_item_list_lookup;
|
2007-01-11 21:18:01 +01:00
|
|
|
thd->lex->current_select->cur_pos_in_select_list= UNDEF_POS;
|
|
|
|
|
2005-10-15 23:32:37 +02:00
|
|
|
thd->lex->allow_sum_func= save_allow_sum_func;
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
thd->mark_used_columns= save_mark_used_columns;
|
|
|
|
DBUG_PRINT("info", ("thd->mark_used_columns: %d", thd->mark_used_columns));
|
2007-10-30 18:08:16 +01:00
|
|
|
DBUG_RETURN(test(thd->is_error()));
|
2003-01-25 01:25:52 +01:00
|
|
|
}
|
2001-03-21 19:13:46 +01:00
|
|
|
|
2003-08-23 12:29:38 +02:00
|
|
|
|
2004-09-14 18:28:29 +02:00
|
|
|
/*
|
|
|
|
make list of leaves of join table tree
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
make_leaves_list()
|
|
|
|
list pointer to pointer on list first element
|
|
|
|
tables table list
|
|
|
|
|
|
|
|
RETURN pointer on pointer to next_leaf of last element
|
|
|
|
*/
|
|
|
|
|
|
|
|
TABLE_LIST **make_leaves_list(TABLE_LIST **list, TABLE_LIST *tables)
|
|
|
|
{
|
|
|
|
for (TABLE_LIST *table= tables; table; table= table->next_local)
|
|
|
|
{
|
2005-10-27 23:18:23 +02:00
|
|
|
if (table->merge_underlying_list)
|
|
|
|
{
|
|
|
|
DBUG_ASSERT(table->view &&
|
|
|
|
table->effective_algorithm == VIEW_ALGORITHM_MERGE);
|
|
|
|
list= make_leaves_list(list, table->merge_underlying_list);
|
|
|
|
}
|
2004-09-14 18:28:29 +02:00
|
|
|
else
|
|
|
|
{
|
|
|
|
*list= table;
|
|
|
|
list= &table->next_leaf;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return list;
|
|
|
|
}
|
|
|
|
|
2001-03-21 19:13:46 +01:00
|
|
|
/*
|
2004-06-25 15:52:01 +02:00
|
|
|
prepare tables
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
setup_tables()
|
2004-11-25 01:23:13 +01:00
|
|
|
thd Thread handler
|
2005-07-01 06:05:42 +02:00
|
|
|
context name resolution contest to setup table list there
|
2005-08-12 16:57:19 +02:00
|
|
|
from_clause Top-level list of table references in the FROM clause
|
|
|
|
tables Table list (select_lex->table_list)
|
|
|
|
leaves List of join table leaves list (select_lex->leaf_tables)
|
2004-11-25 01:23:13 +01:00
|
|
|
refresh It is onle refresh for subquery
|
|
|
|
select_insert It is SELECT ... INSERT command
|
2004-06-25 15:52:01 +02:00
|
|
|
|
2004-10-29 18:26:52 +02:00
|
|
|
NOTE
|
|
|
|
Check also that the 'used keys' and 'ignored keys' exists and set up the
|
2005-08-12 16:57:19 +02:00
|
|
|
table structure accordingly.
|
|
|
|
Create a list of leaf tables. For queries with NATURAL/USING JOINs,
|
|
|
|
compute the row types of the top most natural/using join table references
|
|
|
|
and link these into a list of table references for name resolution.
|
2003-02-12 20:55:37 +01:00
|
|
|
|
2004-10-29 18:26:52 +02:00
|
|
|
This has to be called for all tables that are used by items, as otherwise
|
|
|
|
table->map is not set and all Item_field will be regarded as const items.
|
2004-07-16 00:15:55 +02:00
|
|
|
|
2004-10-29 18:26:52 +02:00
|
|
|
RETURN
|
2005-08-12 16:57:19 +02:00
|
|
|
FALSE ok; In this case *map will includes the chosen index
|
2004-11-25 01:23:13 +01:00
|
|
|
TRUE error
|
2001-03-21 19:13:46 +01:00
|
|
|
*/
|
|
|
|
|
2005-07-01 06:05:42 +02:00
|
|
|
bool setup_tables(THD *thd, Name_resolution_context *context,
|
2005-08-12 16:57:19 +02:00
|
|
|
List<TABLE_LIST> *from_clause, TABLE_LIST *tables,
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
TABLE_LIST **leaves, bool select_insert)
|
2001-03-21 19:13:46 +01:00
|
|
|
{
|
2004-12-30 23:44:00 +01:00
|
|
|
uint tablenr= 0;
|
2001-03-21 19:13:46 +01:00
|
|
|
DBUG_ENTER("setup_tables");
|
2005-07-01 06:05:42 +02:00
|
|
|
|
2007-02-19 13:39:37 +01:00
|
|
|
DBUG_ASSERT ((select_insert && !tables->next_name_resolution_table) || !tables ||
|
|
|
|
(context->table_list && context->first_name_resolution_table));
|
2004-11-25 01:23:13 +01:00
|
|
|
/*
|
|
|
|
this is used for INSERT ... SELECT.
|
2004-11-25 08:28:32 +01:00
|
|
|
For select we setup tables except first (and its underlying tables)
|
2004-11-25 01:23:13 +01:00
|
|
|
*/
|
|
|
|
TABLE_LIST *first_select_table= (select_insert ?
|
|
|
|
tables->next_local:
|
|
|
|
0);
|
2004-09-14 18:28:29 +02:00
|
|
|
if (!(*leaves))
|
|
|
|
make_leaves_list(leaves, tables);
|
|
|
|
|
2005-06-02 15:32:02 +02:00
|
|
|
TABLE_LIST *table_list;
|
|
|
|
for (table_list= *leaves;
|
2004-07-16 00:15:55 +02:00
|
|
|
table_list;
|
2004-09-14 18:28:29 +02:00
|
|
|
table_list= table_list->next_leaf, tablenr++)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2003-07-02 00:45:22 +02:00
|
|
|
TABLE *table= table_list->table;
|
2006-06-01 07:55:45 +02:00
|
|
|
table->pos_in_table_list= table_list;
|
2004-11-25 01:23:13 +01:00
|
|
|
if (first_select_table &&
|
2005-08-02 21:54:49 +02:00
|
|
|
table_list->top_table() == first_select_table)
|
2004-11-25 01:23:13 +01:00
|
|
|
{
|
|
|
|
/* new counting for SELECT of INSERT ... SELECT command */
|
|
|
|
first_select_table= 0;
|
|
|
|
tablenr= 0;
|
|
|
|
}
|
2003-07-02 00:45:22 +02:00
|
|
|
setup_table_map(table, table_list, tablenr);
|
2007-03-05 18:08:41 +01:00
|
|
|
if (table_list->process_index_hints(table))
|
|
|
|
DBUG_RETURN(1);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
if (tablenr > MAX_TABLES)
|
|
|
|
{
|
|
|
|
my_error(ER_TOO_MANY_TABLES,MYF(0),MAX_TABLES);
|
2001-03-21 19:13:46 +01:00
|
|
|
DBUG_RETURN(1);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2005-06-02 15:32:02 +02:00
|
|
|
for (table_list= tables;
|
2005-05-30 18:54:37 +02:00
|
|
|
table_list;
|
|
|
|
table_list= table_list->next_local)
|
2004-09-14 18:28:29 +02:00
|
|
|
{
|
2005-10-27 23:18:23 +02:00
|
|
|
if (table_list->merge_underlying_list)
|
2005-07-01 06:05:42 +02:00
|
|
|
{
|
2005-10-27 23:18:23 +02:00
|
|
|
DBUG_ASSERT(table_list->view &&
|
|
|
|
table_list->effective_algorithm == VIEW_ALGORITHM_MERGE);
|
2005-09-02 15:21:19 +02:00
|
|
|
Query_arena *arena= thd->stmt_arena, backup;
|
2005-07-01 06:05:42 +02:00
|
|
|
bool res;
|
|
|
|
if (arena->is_conventional())
|
|
|
|
arena= 0; // For easier test
|
|
|
|
else
|
2005-09-02 15:21:19 +02:00
|
|
|
thd->set_n_backup_active_arena(arena, &backup);
|
2005-10-27 23:18:23 +02:00
|
|
|
res= table_list->setup_underlying(thd);
|
2005-07-01 06:05:42 +02:00
|
|
|
if (arena)
|
2005-09-02 15:21:19 +02:00
|
|
|
thd->restore_active_arena(arena, &backup);
|
2005-07-01 06:05:42 +02:00
|
|
|
if (res)
|
|
|
|
DBUG_RETURN(1);
|
|
|
|
}
|
2004-09-14 18:28:29 +02:00
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
/* Precompute and store the row types of NATURAL/USING joins. */
|
|
|
|
if (setup_natural_join_row_types(thd, from_clause, context))
|
|
|
|
DBUG_RETURN(1);
|
|
|
|
|
2001-03-21 19:13:46 +01:00
|
|
|
DBUG_RETURN(0);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2000-08-28 15:43:58 +02:00
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2006-05-26 10:47:53 +02:00
|
|
|
/*
|
|
|
|
prepare tables and check access for the view tables
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
setup_tables_and_check_view_access()
|
|
|
|
thd Thread handler
|
|
|
|
context name resolution contest to setup table list there
|
|
|
|
from_clause Top-level list of table references in the FROM clause
|
|
|
|
tables Table list (select_lex->table_list)
|
|
|
|
conds Condition of current SELECT (can be changed by VIEW)
|
|
|
|
leaves List of join table leaves list (select_lex->leaf_tables)
|
|
|
|
refresh It is onle refresh for subquery
|
|
|
|
select_insert It is SELECT ... INSERT command
|
|
|
|
want_access what access is needed
|
|
|
|
|
|
|
|
NOTE
|
|
|
|
a wrapper for check_tables that will also check the resulting
|
|
|
|
table leaves list for access to all the tables that belong to a view
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
FALSE ok; In this case *map will include the chosen index
|
|
|
|
TRUE error
|
|
|
|
*/
|
|
|
|
bool setup_tables_and_check_access(THD *thd,
|
|
|
|
Name_resolution_context *context,
|
|
|
|
List<TABLE_LIST> *from_clause,
|
|
|
|
TABLE_LIST *tables,
|
2006-06-05 05:16:08 +02:00
|
|
|
TABLE_LIST **leaves,
|
2006-05-26 10:47:53 +02:00
|
|
|
bool select_insert,
|
2006-08-30 20:11:23 +02:00
|
|
|
ulong want_access_first,
|
2006-05-26 10:47:53 +02:00
|
|
|
ulong want_access)
|
|
|
|
{
|
2006-06-05 05:16:08 +02:00
|
|
|
TABLE_LIST *leaves_tmp= NULL;
|
2006-08-15 19:45:24 +02:00
|
|
|
bool first_table= true;
|
2006-05-26 10:47:53 +02:00
|
|
|
|
2006-06-05 05:16:08 +02:00
|
|
|
if (setup_tables(thd, context, from_clause, tables,
|
|
|
|
&leaves_tmp, select_insert))
|
2006-05-26 10:47:53 +02:00
|
|
|
return TRUE;
|
|
|
|
|
2007-03-22 09:19:41 +01:00
|
|
|
if (leaves)
|
|
|
|
*leaves= leaves_tmp;
|
2006-05-26 10:47:53 +02:00
|
|
|
|
|
|
|
for (; leaves_tmp; leaves_tmp= leaves_tmp->next_leaf)
|
2006-06-05 05:16:08 +02:00
|
|
|
{
|
2006-05-26 10:47:53 +02:00
|
|
|
if (leaves_tmp->belong_to_view &&
|
2006-08-15 19:45:24 +02:00
|
|
|
check_single_table_access(thd, first_table ? want_access_first :
|
2007-04-04 11:01:47 +02:00
|
|
|
want_access, leaves_tmp, FALSE))
|
2006-05-26 10:47:53 +02:00
|
|
|
{
|
|
|
|
tables->hide_view_error(thd);
|
|
|
|
return TRUE;
|
|
|
|
}
|
2006-08-29 16:58:50 +02:00
|
|
|
first_table= 0;
|
2006-06-05 05:16:08 +02:00
|
|
|
}
|
2006-05-26 10:47:53 +02:00
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2003-11-18 12:47:27 +01:00
|
|
|
/*
|
|
|
|
Create a key_map from a list of index names
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
get_key_map_from_key_list()
|
|
|
|
map key_map to fill in
|
|
|
|
table Table
|
|
|
|
index_list List of index names
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
0 ok; In this case *map will includes the choosed index
|
|
|
|
1 error
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool get_key_map_from_key_list(key_map *map, TABLE *table,
|
2003-10-11 13:06:55 +02:00
|
|
|
List<String> *index_list)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2001-08-02 05:29:50 +02:00
|
|
|
List_iterator_fast<String> it(*index_list);
|
2000-07-31 21:29:14 +02:00
|
|
|
String *name;
|
|
|
|
uint pos;
|
2003-10-11 13:06:55 +02:00
|
|
|
|
|
|
|
map->clear_all();
|
2000-07-31 21:29:14 +02:00
|
|
|
while ((name=it++))
|
|
|
|
{
|
2005-01-06 12:00:13 +01:00
|
|
|
if (table->s->keynames.type_names == 0 ||
|
|
|
|
(pos= find_type(&table->s->keynames, name->ptr(),
|
|
|
|
name->length(), 1)) <=
|
2004-07-22 13:05:00 +02:00
|
|
|
0)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2006-06-06 19:21:36 +02:00
|
|
|
my_error(ER_KEY_DOES_NOT_EXITS, MYF(0), name->c_ptr(),
|
2006-06-01 07:55:45 +02:00
|
|
|
table->pos_in_table_list->alias);
|
2003-10-11 13:06:55 +02:00
|
|
|
map->set_all();
|
2003-11-18 12:47:27 +01:00
|
|
|
return 1;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2003-10-11 13:06:55 +02:00
|
|
|
map->set_bit(pos-1);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2003-11-18 12:47:27 +01:00
|
|
|
return 0;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2003-11-18 12:47:27 +01:00
|
|
|
|
2004-09-03 20:43:04 +02:00
|
|
|
/*
|
|
|
|
Drops in all fields instead of current '*' field
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
insert_fields()
|
|
|
|
thd Thread handler
|
2005-07-01 06:05:42 +02:00
|
|
|
context Context for name resolution
|
2004-09-03 20:43:04 +02:00
|
|
|
db_name Database name in case of 'database_name.table_name.*'
|
|
|
|
table_name Table name in case of 'table_name.*'
|
|
|
|
it Pointer to '*'
|
|
|
|
any_privileges 0 If we should ensure that we have SELECT privileges
|
|
|
|
for all columns
|
|
|
|
1 If any privilege is ok
|
|
|
|
RETURN
|
2005-08-17 12:46:36 +02:00
|
|
|
0 ok 'it' is updated to point at last inserted
|
2004-11-03 11:39:38 +01:00
|
|
|
1 error. Error message is generated but not sent to client
|
2004-09-03 20:43:04 +02:00
|
|
|
*/
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2001-07-08 03:15:41 +02:00
|
|
|
bool
|
2005-07-01 06:05:42 +02:00
|
|
|
insert_fields(THD *thd, Name_resolution_context *context, const char *db_name,
|
2004-07-16 00:15:55 +02:00
|
|
|
const char *table_name, List_iterator<Item> *it,
|
2005-06-21 19:15:21 +02:00
|
|
|
bool any_privileges)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
Field_iterator_table_ref field_iterator;
|
|
|
|
bool found;
|
2004-09-06 14:14:10 +02:00
|
|
|
char name_buff[NAME_LEN+1];
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("insert_fields");
|
2005-09-02 15:21:19 +02:00
|
|
|
DBUG_PRINT("arena", ("stmt arena: 0x%lx", (ulong)thd->stmt_arena));
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2004-08-25 15:12:31 +02:00
|
|
|
if (db_name && lower_case_table_names)
|
|
|
|
{
|
2004-08-27 23:49:54 +02:00
|
|
|
/*
|
|
|
|
convert database to lower case for comparison
|
|
|
|
We can't do this in Item_field as this would change the
|
|
|
|
'name' of the item which may be used in the select list
|
|
|
|
*/
|
2004-08-26 18:22:22 +02:00
|
|
|
strmake(name_buff, db_name, sizeof(name_buff)-1);
|
2004-09-02 15:57:26 +02:00
|
|
|
my_casedn_str(files_charset_info, name_buff);
|
2004-08-27 23:49:54 +02:00
|
|
|
db_name= name_buff;
|
2004-08-25 15:12:31 +02:00
|
|
|
}
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
found= FALSE;
|
2005-08-23 17:08:04 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
If table names are qualified, then loop over all tables used in the query,
|
|
|
|
else treat natural joins as leaves and do not iterate over their underlying
|
|
|
|
tables.
|
|
|
|
*/
|
|
|
|
for (TABLE_LIST *tables= (table_name ? context->table_list :
|
|
|
|
context->first_name_resolution_table);
|
2005-07-01 06:05:42 +02:00
|
|
|
tables;
|
2005-08-23 17:08:04 +02:00
|
|
|
tables= (table_name ? tables->next_local :
|
|
|
|
tables->next_name_resolution_table)
|
|
|
|
)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2004-09-03 20:43:04 +02:00
|
|
|
Field *field;
|
|
|
|
TABLE *table= tables->table;
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_ASSERT(tables->is_leaf_for_name_resolution());
|
|
|
|
|
2009-06-10 16:04:07 +02:00
|
|
|
if ((table_name && my_strcasecmp(table_alias_charset, table_name,
|
|
|
|
tables->alias)) ||
|
2005-08-23 17:08:04 +02:00
|
|
|
(db_name && strcmp(tables->db,db_name)))
|
|
|
|
continue;
|
2005-08-12 16:57:19 +02:00
|
|
|
|
2003-09-26 12:33:13 +02:00
|
|
|
#ifndef NO_EMBEDDED_ACCESS_CHECKS
|
2008-09-03 16:45:40 +02:00
|
|
|
/*
|
|
|
|
Ensure that we have access rights to all fields to be inserted. Under
|
|
|
|
some circumstances, this check may be skipped.
|
|
|
|
|
|
|
|
- If any_privileges is true, skip the check.
|
|
|
|
|
|
|
|
- If the SELECT privilege has been found as fulfilled already for both
|
|
|
|
the TABLE and TABLE_LIST objects (and both of these exist, of
|
|
|
|
course), the check is skipped.
|
|
|
|
|
|
|
|
- If the SELECT privilege has been found fulfilled for the TABLE object
|
|
|
|
and the TABLE_LIST represents a derived table other than a view (see
|
|
|
|
below), the check is skipped.
|
|
|
|
|
|
|
|
- If the TABLE_LIST object represents a view, we may skip checking if
|
|
|
|
the SELECT privilege has been found fulfilled for it, regardless of
|
|
|
|
the TABLE object.
|
|
|
|
|
|
|
|
- If there is no TABLE object, the test is skipped if either
|
|
|
|
* the TABLE_LIST does not represent a view, or
|
|
|
|
* the SELECT privilege has been found fulfilled.
|
|
|
|
|
|
|
|
A TABLE_LIST that is not a view may be a subquery, an
|
|
|
|
information_schema table, or a nested table reference. See the comment
|
|
|
|
for TABLE_LIST.
|
|
|
|
*/
|
2009-06-10 16:04:07 +02:00
|
|
|
if (!((table && !tables->view && (table->grant.privilege & SELECT_ACL)) ||
|
2009-06-17 16:56:44 +02:00
|
|
|
(tables->view && (tables->grant.privilege & SELECT_ACL))) &&
|
2005-08-12 16:57:19 +02:00
|
|
|
!any_privileges)
|
|
|
|
{
|
|
|
|
field_iterator.set(tables);
|
2007-09-27 11:15:19 +02:00
|
|
|
if (check_grant_all_columns(thd, SELECT_ACL, &field_iterator))
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN(TRUE);
|
|
|
|
}
|
2003-09-26 12:33:13 +02:00
|
|
|
#endif
|
2004-07-16 00:15:55 +02:00
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
/*
|
|
|
|
Update the tables used in the query based on the referenced fields. For
|
|
|
|
views and natural joins this update is performed inside the loop below.
|
|
|
|
*/
|
|
|
|
if (table)
|
|
|
|
thd->used_tables|= table->map;
|
|
|
|
|
|
|
|
/*
|
|
|
|
Initialize a generic field iterator for the current table reference.
|
|
|
|
Notice that it is guaranteed that this iterator will iterate over the
|
|
|
|
fields of a single table reference, because 'tables' is a leaf (for
|
|
|
|
name resolution purposes).
|
|
|
|
*/
|
|
|
|
field_iterator.set(tables);
|
|
|
|
|
|
|
|
for (; !field_iterator.end_of_fields(); field_iterator.next())
|
|
|
|
{
|
|
|
|
Item *item;
|
|
|
|
|
|
|
|
if (!(item= field_iterator.create_item(thd)))
|
|
|
|
DBUG_RETURN(TRUE);
|
2008-10-17 16:55:06 +02:00
|
|
|
DBUG_ASSERT(item->fixed);
|
|
|
|
/* cache the table for the Item_fields inserted by expanding stars */
|
|
|
|
if (item->type() == Item::FIELD_ITEM && tables->cacheable_table)
|
|
|
|
((Item_field *)item)->cached_table= tables;
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
if (!found)
|
2004-09-14 18:28:29 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
found= TRUE;
|
2005-08-17 12:46:36 +02:00
|
|
|
it->replace(item); /* Replace '*' with the first found item. */
|
2004-09-14 18:28:29 +02:00
|
|
|
}
|
2004-07-16 00:15:55 +02:00
|
|
|
else
|
2005-08-12 16:57:19 +02:00
|
|
|
it->after(item); /* Add 'item' to the SELECT list. */
|
|
|
|
|
|
|
|
#ifndef NO_EMBEDDED_ACCESS_CHECKS
|
|
|
|
/*
|
|
|
|
Set privilege information for the fields of newly created views.
|
|
|
|
We have that (any_priviliges == TRUE) if and only if we are creating
|
|
|
|
a view. In the time of view creation we can't use the MERGE algorithm,
|
2005-08-17 12:46:36 +02:00
|
|
|
therefore if 'tables' is itself a view, it is represented by a
|
|
|
|
temporary table. Thus in this case we can be sure that 'item' is an
|
|
|
|
Item_field.
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
|
|
|
if (any_privileges)
|
2004-09-14 18:28:29 +02:00
|
|
|
{
|
2009-06-10 16:04:07 +02:00
|
|
|
DBUG_ASSERT((tables->field_translation == NULL && table) ||
|
2005-08-12 16:57:19 +02:00
|
|
|
tables->is_natural_join);
|
|
|
|
DBUG_ASSERT(item->type() == Item::FIELD_ITEM);
|
|
|
|
Item_field *fld= (Item_field*) item;
|
2008-09-03 16:45:40 +02:00
|
|
|
const char *field_table_name= field_iterator.get_table_name();
|
2005-08-17 12:46:36 +02:00
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
if (!tables->schema_table &&
|
|
|
|
!(fld->have_privileges=
|
|
|
|
(get_column_grant(thd, field_iterator.grant(),
|
2008-09-03 16:45:40 +02:00
|
|
|
field_iterator.get_db_name(),
|
2005-08-23 17:08:04 +02:00
|
|
|
field_table_name, fld->field_name) &
|
2005-08-12 16:57:19 +02:00
|
|
|
VIEW_ANY_ACL)))
|
|
|
|
{
|
2008-09-03 16:45:40 +02:00
|
|
|
my_error(ER_TABLEACCESS_DENIED_ERROR, MYF(0), "ANY",
|
2005-09-15 21:29:07 +02:00
|
|
|
thd->security_ctx->priv_user,
|
|
|
|
thd->security_ctx->host_or_ip,
|
2008-09-03 16:45:40 +02:00
|
|
|
field_table_name);
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN(TRUE);
|
|
|
|
}
|
2004-09-14 18:28:29 +02:00
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
#endif
|
2005-07-01 06:05:42 +02:00
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
if ((field= field_iterator.field()))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
/* Mark fields as used to allow storage engine to optimze access */
|
|
|
|
bitmap_set_bit(field->table->read_set, field->field_index);
|
2005-08-12 16:57:19 +02:00
|
|
|
if (table)
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
{
|
2007-03-05 18:08:41 +01:00
|
|
|
table->covering_keys.intersect(field->part_of_key);
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
table->merge_keys.merge(field->part_of_key);
|
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
if (tables->is_natural_join)
|
2004-06-25 15:52:01 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
TABLE *field_table;
|
|
|
|
/*
|
2005-08-17 12:46:36 +02:00
|
|
|
In this case we are sure that the column ref will not be created
|
2005-08-12 16:57:19 +02:00
|
|
|
because it was already created and stored with the natural join.
|
2005-08-17 12:46:36 +02:00
|
|
|
*/
|
2005-08-12 16:57:19 +02:00
|
|
|
Natural_join_column *nj_col;
|
2006-03-02 10:50:15 +01:00
|
|
|
if (!(nj_col= field_iterator.get_natural_column_ref()))
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN(TRUE);
|
2006-03-02 10:50:15 +01:00
|
|
|
DBUG_ASSERT(nj_col->table_field);
|
2005-08-12 16:57:19 +02:00
|
|
|
field_table= nj_col->table_ref->table;
|
|
|
|
if (field_table)
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
thd->used_tables|= field_table->map;
|
2007-03-05 18:08:41 +01:00
|
|
|
field_table->covering_keys.intersect(field->part_of_key);
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
field_table->merge_keys.merge(field->part_of_key);
|
2005-08-12 16:57:19 +02:00
|
|
|
field_table->used_fields++;
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
2005-07-12 14:18:05 +02:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
else
|
|
|
|
thd->used_tables|= item->used_tables();
|
2007-04-24 20:35:57 +02:00
|
|
|
thd->lex->current_select->cur_pos_in_select_list++;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2005-08-12 16:57:19 +02:00
|
|
|
/*
|
|
|
|
In case of stored tables, all fields are considered as used,
|
|
|
|
while in the case of views, the fields considered as used are the
|
|
|
|
ones marked in setup_tables during fix_fields of view columns.
|
|
|
|
For NATURAL joins, used_tables is updated in the IF above.
|
|
|
|
*/
|
|
|
|
if (table)
|
|
|
|
table->used_fields= table->s->fields;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2004-09-03 20:43:04 +02:00
|
|
|
if (found)
|
2005-08-12 16:57:19 +02:00
|
|
|
DBUG_RETURN(FALSE);
|
2004-09-03 20:43:04 +02:00
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
/*
|
2005-08-17 12:46:36 +02:00
|
|
|
TODO: in the case when we skipped all columns because there was a
|
|
|
|
qualified '*', and all columns were coalesced, we have to give a more
|
|
|
|
meaningful message than ER_BAD_TABLE_ERROR.
|
2005-08-12 16:57:19 +02:00
|
|
|
*/
|
2004-09-03 20:43:04 +02:00
|
|
|
if (!table_name)
|
2004-11-12 13:34:00 +01:00
|
|
|
my_message(ER_NO_TABLES_USED, ER(ER_NO_TABLES_USED), MYF(0));
|
2004-09-03 20:43:04 +02:00
|
|
|
else
|
|
|
|
my_error(ER_BAD_TABLE_ERROR, MYF(0), table_name);
|
2005-08-12 16:57:19 +02:00
|
|
|
|
|
|
|
DBUG_RETURN(TRUE);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
2005-08-12 16:57:19 +02:00
|
|
|
Fix all conditions and outer join expressions.
|
2004-09-14 18:28:29 +02:00
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
setup_conds()
|
|
|
|
thd thread handler
|
2005-08-12 16:57:19 +02:00
|
|
|
tables list of tables for name resolving (select_lex->table_list)
|
|
|
|
leaves list of leaves of join table tree (select_lex->leaf_tables)
|
|
|
|
conds WHERE clause
|
|
|
|
|
|
|
|
DESCRIPTION
|
|
|
|
TODO
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
TRUE if some error occured (e.g. out of memory)
|
|
|
|
FALSE if all is OK
|
2000-07-31 21:29:14 +02:00
|
|
|
*/
|
|
|
|
|
2005-07-01 06:05:42 +02:00
|
|
|
int setup_conds(THD *thd, TABLE_LIST *tables, TABLE_LIST *leaves,
|
|
|
|
COND **conds)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2004-06-25 15:52:01 +02:00
|
|
|
SELECT_LEX *select_lex= thd->lex->current_select;
|
2005-09-02 15:21:19 +02:00
|
|
|
Query_arena *arena= thd->stmt_arena, backup;
|
2004-11-12 12:28:34 +01:00
|
|
|
TABLE_LIST *table= NULL; // For HP compilers
|
2005-07-01 06:05:42 +02:00
|
|
|
/*
|
|
|
|
it_is_update set to TRUE when tables of primary SELECT_LEX (SELECT_LEX
|
|
|
|
which belong to LEX, i.e. most up SELECT) will be updated by
|
|
|
|
INSERT/UPDATE/LOAD
|
|
|
|
NOTE: using this condition helps to prevent call of prepare_check_option()
|
|
|
|
from subquery of VIEW, because tables of subquery belongs to VIEW
|
|
|
|
(see condition before prepare_check_option() call)
|
|
|
|
*/
|
|
|
|
bool it_is_update= (select_lex == &thd->lex->select_lex) &&
|
|
|
|
thd->lex->which_check_option_applicable();
|
2007-03-05 21:33:57 +01:00
|
|
|
bool save_is_item_list_lookup= select_lex->is_item_list_lookup;
|
|
|
|
select_lex->is_item_list_lookup= 0;
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("setup_conds");
|
2004-06-25 15:52:01 +02:00
|
|
|
|
2004-09-06 14:14:10 +02:00
|
|
|
if (select_lex->conds_processed_with_permanent_arena ||
|
2004-09-10 01:22:44 +02:00
|
|
|
arena->is_conventional())
|
2004-09-06 14:14:10 +02:00
|
|
|
arena= 0; // For easier test
|
|
|
|
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
thd->mark_used_columns= MARK_COLUMNS_READ;
|
|
|
|
DBUG_PRINT("info", ("thd->mark_used_columns: %d", thd->mark_used_columns));
|
2005-08-12 10:27:04 +02:00
|
|
|
select_lex->cond_count= 0;
|
2006-10-16 23:25:28 +02:00
|
|
|
select_lex->between_count= 0;
|
2007-08-15 19:24:18 +02:00
|
|
|
select_lex->max_equal_elems= 0;
|
2004-07-16 00:15:55 +02:00
|
|
|
|
2005-07-01 06:05:42 +02:00
|
|
|
for (table= tables; table; table= table->next_local)
|
|
|
|
{
|
|
|
|
if (table->prepare_where(thd, conds, FALSE))
|
|
|
|
goto err_no_arena;
|
|
|
|
}
|
|
|
|
|
2000-07-31 21:29:14 +02:00
|
|
|
if (*conds)
|
|
|
|
{
|
|
|
|
thd->where="where clause";
|
2009-06-10 16:04:07 +02:00
|
|
|
if ((!(*conds)->fixed && (*conds)->fix_fields(thd, conds)) ||
|
2004-06-25 15:52:01 +02:00
|
|
|
(*conds)->check_cols(1))
|
2004-10-10 10:01:05 +02:00
|
|
|
goto err_no_arena;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
2005-08-12 16:57:19 +02:00
|
|
|
/*
|
|
|
|
Apply fix_fields() to all ON clauses at all levels of nesting,
|
|
|
|
including the ones inside view definitions.
|
|
|
|
*/
|
2004-11-21 19:08:12 +01:00
|
|
|
for (table= leaves; table; table= table->next_leaf)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2005-08-12 16:57:19 +02:00
|
|
|
TABLE_LIST *embedded; /* The table at the current level of nesting. */
|
|
|
|
TABLE_LIST *embedding= table; /* The parent nested table reference. */
|
2004-06-11 07:27:21 +02:00
|
|
|
do
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2004-06-11 07:27:21 +02:00
|
|
|
embedded= embedding;
|
|
|
|
if (embedded->on_expr)
|
2000-09-25 23:33:25 +02:00
|
|
|
{
|
2004-06-11 07:27:21 +02:00
|
|
|
/* Make a join an a expression */
|
|
|
|
thd->where="on clause";
|
2009-06-10 16:04:07 +02:00
|
|
|
if ((!embedded->on_expr->fixed &&
|
|
|
|
embedded->on_expr->fix_fields(thd, &embedded->on_expr)) ||
|
2004-06-11 07:27:21 +02:00
|
|
|
embedded->on_expr->check_cols(1))
|
2004-10-10 10:01:05 +02:00
|
|
|
goto err_no_arena;
|
2004-06-25 15:52:01 +02:00
|
|
|
select_lex->cond_count++;
|
2000-09-25 23:33:25 +02:00
|
|
|
}
|
2004-06-11 07:27:21 +02:00
|
|
|
embedding= embedded->embedding;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2004-06-11 07:27:21 +02:00
|
|
|
while (embedding &&
|
|
|
|
embedding->nested_join->join_list.head() == embedded);
|
2005-07-01 06:05:42 +02:00
|
|
|
|
|
|
|
/* process CHECK OPTION */
|
|
|
|
if (it_is_update)
|
|
|
|
{
|
2005-09-02 08:50:17 +02:00
|
|
|
TABLE_LIST *view= table->top_table();
|
2005-07-01 06:05:42 +02:00
|
|
|
if (view->effective_with_check)
|
|
|
|
{
|
|
|
|
if (view->prepare_check_option(thd))
|
|
|
|
goto err_no_arena;
|
|
|
|
thd->change_item_tree(&table->check_option, view->check_option);
|
|
|
|
}
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2004-06-25 15:52:01 +02:00
|
|
|
|
2005-09-02 15:21:19 +02:00
|
|
|
if (!thd->stmt_arena->is_conventional())
|
2004-06-25 15:52:01 +02:00
|
|
|
{
|
|
|
|
/*
|
|
|
|
We are in prepared statement preparation code => we should store
|
|
|
|
WHERE clause changing for next executions.
|
|
|
|
|
|
|
|
We do this ON -> WHERE transformation only once per PS/SP statement.
|
|
|
|
*/
|
|
|
|
select_lex->where= *conds;
|
|
|
|
select_lex->conds_processed_with_permanent_arena= 1;
|
|
|
|
}
|
2007-03-03 22:47:42 +01:00
|
|
|
thd->lex->current_select->is_item_list_lookup= save_is_item_list_lookup;
|
2007-10-30 18:08:16 +01:00
|
|
|
DBUG_RETURN(test(thd->is_error()));
|
2004-06-25 15:52:01 +02:00
|
|
|
|
2004-10-10 10:01:05 +02:00
|
|
|
err_no_arena:
|
2007-03-05 21:33:57 +01:00
|
|
|
select_lex->is_item_list_lookup= save_is_item_list_lookup;
|
2004-06-25 15:52:01 +02:00
|
|
|
DBUG_RETURN(1);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/******************************************************************************
|
|
|
|
** Fill a record with data (for INSERT or UPDATE)
|
|
|
|
** Returns : 1 if some field has wrong type
|
|
|
|
******************************************************************************/
|
|
|
|
|
2004-11-12 13:34:00 +01:00
|
|
|
|
|
|
|
/*
|
|
|
|
Fill fields with given items.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
fill_record()
|
|
|
|
thd thread handler
|
|
|
|
fields Item_fields list to be filled
|
|
|
|
values values to fill with
|
|
|
|
ignore_errors TRUE if we should ignore errors
|
|
|
|
|
2007-03-30 16:13:33 +02:00
|
|
|
NOTE
|
|
|
|
fill_record() may set table->auto_increment_field_not_null and a
|
|
|
|
caller should make sure that it is reset after their last call to this
|
|
|
|
function.
|
|
|
|
|
2004-11-12 13:34:00 +01:00
|
|
|
RETURN
|
|
|
|
FALSE OK
|
|
|
|
TRUE error occured
|
|
|
|
*/
|
|
|
|
|
2005-05-24 20:19:33 +02:00
|
|
|
static bool
|
2004-11-12 13:34:00 +01:00
|
|
|
fill_record(THD * thd, List<Item> &fields, List<Item> &values,
|
|
|
|
bool ignore_errors)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2001-08-02 05:29:50 +02:00
|
|
|
List_iterator_fast<Item> f(fields),v(values);
|
2006-07-25 17:42:49 +02:00
|
|
|
Item *value, *fld;
|
2000-07-31 21:29:14 +02:00
|
|
|
Item_field *field;
|
2007-03-30 16:13:33 +02:00
|
|
|
TABLE *table= 0;
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("fill_record");
|
|
|
|
|
2007-03-30 16:13:33 +02:00
|
|
|
/*
|
|
|
|
Reset the table->auto_increment_field_not_null as it is valid for
|
|
|
|
only one row.
|
|
|
|
*/
|
|
|
|
if (fields.elements)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
On INSERT or UPDATE fields are checked to be from the same table,
|
|
|
|
thus we safely can take table from the first field.
|
|
|
|
*/
|
|
|
|
fld= (Item_field*)f++;
|
|
|
|
if (!(field= fld->filed_for_view_update()))
|
|
|
|
{
|
|
|
|
my_error(ER_NONUPDATEABLE_COLUMN, MYF(0), fld->name);
|
|
|
|
goto err;
|
|
|
|
}
|
|
|
|
table= field->field->table;
|
|
|
|
table->auto_increment_field_not_null= FALSE;
|
|
|
|
f.rewind();
|
|
|
|
}
|
2006-07-25 17:42:49 +02:00
|
|
|
while ((fld= f++))
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2006-07-25 17:42:49 +02:00
|
|
|
if (!(field= fld->filed_for_view_update()))
|
|
|
|
{
|
|
|
|
my_error(ER_NONUPDATEABLE_COLUMN, MYF(0), fld->name);
|
2007-03-30 16:13:33 +02:00
|
|
|
goto err;
|
2006-07-25 17:42:49 +02:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
value=v++;
|
2003-07-23 00:58:30 +02:00
|
|
|
Field *rfield= field->field;
|
2007-03-30 16:13:33 +02:00
|
|
|
table= rfield->table;
|
2003-07-28 16:57:46 +02:00
|
|
|
if (rfield == table->next_number_field)
|
2004-07-30 09:47:56 +02:00
|
|
|
table->auto_increment_field_not_null= TRUE;
|
2004-06-25 15:52:01 +02:00
|
|
|
if ((value->save_in_field(rfield, 0) < 0) && !ignore_errors)
|
2004-11-12 13:34:00 +01:00
|
|
|
{
|
|
|
|
my_message(ER_UNKNOWN_ERROR, ER(ER_UNKNOWN_ERROR), MYF(0));
|
2007-03-30 16:13:33 +02:00
|
|
|
goto err;
|
2004-11-12 13:34:00 +01:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2007-10-30 18:08:16 +01:00
|
|
|
DBUG_RETURN(thd->is_error());
|
2007-03-30 16:13:33 +02:00
|
|
|
err:
|
|
|
|
if (table)
|
|
|
|
table->auto_increment_field_not_null= FALSE;
|
|
|
|
DBUG_RETURN(TRUE);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-05-24 20:19:33 +02:00
|
|
|
/*
|
|
|
|
Fill fields in list with values from the list of items and invoke
|
|
|
|
before triggers.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
fill_record_n_invoke_before_triggers()
|
|
|
|
thd thread context
|
|
|
|
fields Item_fields list to be filled
|
|
|
|
values values to fill with
|
|
|
|
ignore_errors TRUE if we should ignore errors
|
|
|
|
triggers object holding list of triggers to be invoked
|
|
|
|
event event type for triggers to be invoked
|
|
|
|
|
|
|
|
NOTE
|
|
|
|
This function assumes that fields which values will be set and triggers
|
|
|
|
to be invoked belong to the same table, and that TABLE::record[0] and
|
|
|
|
record[1] buffers correspond to new and old versions of row respectively.
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
FALSE OK
|
|
|
|
TRUE error occured
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool
|
|
|
|
fill_record_n_invoke_before_triggers(THD *thd, List<Item> &fields,
|
|
|
|
List<Item> &values, bool ignore_errors,
|
|
|
|
Table_triggers_list *triggers,
|
|
|
|
enum trg_event_type event)
|
|
|
|
{
|
|
|
|
return (fill_record(thd, fields, values, ignore_errors) ||
|
2009-06-10 16:04:07 +02:00
|
|
|
(triggers && triggers->process_triggers(thd, event,
|
|
|
|
TRG_ACTION_BEFORE, TRUE)));
|
2005-05-24 20:19:33 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2004-11-12 13:34:00 +01:00
|
|
|
/*
|
|
|
|
Fill field buffer with values from Field list
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
fill_record()
|
|
|
|
thd thread handler
|
|
|
|
ptr pointer on pointer to record
|
|
|
|
values list of fields
|
|
|
|
ignore_errors TRUE if we should ignore errors
|
|
|
|
|
2007-03-30 16:13:33 +02:00
|
|
|
NOTE
|
|
|
|
fill_record() may set table->auto_increment_field_not_null and a
|
|
|
|
caller should make sure that it is reset after their last call to this
|
|
|
|
function.
|
|
|
|
|
2004-11-12 13:34:00 +01:00
|
|
|
RETURN
|
|
|
|
FALSE OK
|
|
|
|
TRUE error occured
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool
|
|
|
|
fill_record(THD *thd, Field **ptr, List<Item> &values, bool ignore_errors)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2001-08-02 05:29:50 +02:00
|
|
|
List_iterator_fast<Item> v(values);
|
2000-07-31 21:29:14 +02:00
|
|
|
Item *value;
|
2007-03-30 16:13:33 +02:00
|
|
|
TABLE *table= 0;
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("fill_record");
|
|
|
|
|
|
|
|
Field *field;
|
2007-03-30 16:13:33 +02:00
|
|
|
/*
|
|
|
|
Reset the table->auto_increment_field_not_null as it is valid for
|
|
|
|
only one row.
|
|
|
|
*/
|
|
|
|
if (*ptr)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
On INSERT or UPDATE fields are checked to be from the same table,
|
|
|
|
thus we safely can take table from the first field.
|
|
|
|
*/
|
|
|
|
table= (*ptr)->table;
|
|
|
|
table->auto_increment_field_not_null= FALSE;
|
|
|
|
}
|
2007-10-30 18:08:16 +01:00
|
|
|
while ((field = *ptr++) && ! thd->is_error())
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
|
|
|
value=v++;
|
2007-03-30 16:13:33 +02:00
|
|
|
table= field->table;
|
2003-07-28 16:57:46 +02:00
|
|
|
if (field == table->next_number_field)
|
2004-07-30 09:47:56 +02:00
|
|
|
table->auto_increment_field_not_null= TRUE;
|
2007-03-23 21:08:31 +01:00
|
|
|
if (value->save_in_field(field, 0) < 0)
|
2007-03-30 16:13:33 +02:00
|
|
|
goto err;
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2007-10-30 18:08:16 +01:00
|
|
|
DBUG_RETURN(thd->is_error());
|
2007-03-30 16:13:33 +02:00
|
|
|
|
|
|
|
err:
|
|
|
|
if (table)
|
|
|
|
table->auto_increment_field_not_null= FALSE;
|
|
|
|
DBUG_RETURN(TRUE);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-05-24 20:19:33 +02:00
|
|
|
/*
|
|
|
|
Fill fields in array with values from the list of items and invoke
|
|
|
|
before triggers.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
fill_record_n_invoke_before_triggers()
|
|
|
|
thd thread context
|
|
|
|
ptr NULL-ended array of fields to be filled
|
|
|
|
values values to fill with
|
|
|
|
ignore_errors TRUE if we should ignore errors
|
|
|
|
triggers object holding list of triggers to be invoked
|
|
|
|
event event type for triggers to be invoked
|
|
|
|
|
|
|
|
NOTE
|
|
|
|
This function assumes that fields which values will be set and triggers
|
|
|
|
to be invoked belong to the same table, and that TABLE::record[0] and
|
|
|
|
record[1] buffers correspond to new and old versions of row respectively.
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
FALSE OK
|
|
|
|
TRUE error occured
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool
|
|
|
|
fill_record_n_invoke_before_triggers(THD *thd, Field **ptr,
|
|
|
|
List<Item> &values, bool ignore_errors,
|
|
|
|
Table_triggers_list *triggers,
|
|
|
|
enum trg_event_type event)
|
|
|
|
{
|
|
|
|
return (fill_record(thd, ptr, values, ignore_errors) ||
|
2009-06-10 16:04:07 +02:00
|
|
|
(triggers && triggers->process_triggers(thd, event,
|
|
|
|
TRG_ACTION_BEFORE, TRUE)));
|
2005-05-24 20:19:33 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-01-10 09:57:03 +01:00
|
|
|
my_bool mysql_rm_tmp_tables(void)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2002-10-08 14:39:37 +02:00
|
|
|
uint i, idx;
|
2007-01-10 09:57:03 +01:00
|
|
|
char filePath[FN_REFLEN], *tmpdir, filePathCopy[FN_REFLEN];
|
2000-07-31 21:29:14 +02:00
|
|
|
MY_DIR *dirp;
|
|
|
|
FILEINFO *file;
|
2007-01-11 13:50:34 +01:00
|
|
|
TABLE_SHARE share;
|
2007-01-10 09:57:03 +01:00
|
|
|
THD *thd;
|
2000-07-31 21:29:14 +02:00
|
|
|
DBUG_ENTER("mysql_rm_tmp_tables");
|
|
|
|
|
2007-01-10 09:57:03 +01:00
|
|
|
if (!(thd= new THD))
|
|
|
|
DBUG_RETURN(1);
|
2007-01-10 11:03:36 +01:00
|
|
|
thd->thread_stack= (char*) &thd;
|
2007-01-10 09:57:03 +01:00
|
|
|
thd->store_globals();
|
|
|
|
|
2002-10-08 14:39:37 +02:00
|
|
|
for (i=0; i<=mysql_tmpdir_list.max; i++)
|
|
|
|
{
|
|
|
|
tmpdir=mysql_tmpdir_list.list[i];
|
2005-11-23 21:45:02 +01:00
|
|
|
/* See if the directory exists */
|
2002-10-08 14:39:37 +02:00
|
|
|
if (!(dirp = my_dir(tmpdir,MYF(MY_WME | MY_DONT_SORT))))
|
|
|
|
continue;
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2002-10-08 14:39:37 +02:00
|
|
|
/* Remove all SQLxxx tables from directory */
|
2000-07-31 21:29:14 +02:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
for (idx=0 ; idx < (uint) dirp->number_off_files ; idx++)
|
|
|
|
{
|
|
|
|
file=dirp->dir_entry+idx;
|
2004-06-25 15:52:01 +02:00
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
/* skiping . and .. */
|
|
|
|
if (file->name[0] == '.' && (!file->name[1] ||
|
|
|
|
(file->name[1] == '.' && !file->name[2])))
|
|
|
|
continue;
|
2004-06-25 15:52:01 +02:00
|
|
|
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
if (!bcmp((uchar*) file->name, (uchar*) tmp_file_prefix,
|
|
|
|
tmp_file_prefix_length))
|
2005-11-23 21:45:02 +01:00
|
|
|
{
|
2007-01-12 10:34:33 +01:00
|
|
|
char *ext= fn_ext(file->name);
|
|
|
|
uint ext_len= strlen(ext);
|
|
|
|
uint filePath_len= my_snprintf(filePath, sizeof(filePath),
|
|
|
|
"%s%c%s", tmpdir, FN_LIBCHAR,
|
|
|
|
file->name);
|
WL#3817: Simplify string / memory area types and make things more consistent (first part)
The following type conversions was done:
- Changed byte to uchar
- Changed gptr to uchar*
- Change my_string to char *
- Change my_size_t to size_t
- Change size_s to size_t
Removed declaration of byte, gptr, my_string, my_size_t and size_s.
Following function parameter changes was done:
- All string functions in mysys/strings was changed to use size_t
instead of uint for string lengths.
- All read()/write() functions changed to use size_t (including vio).
- All protocoll functions changed to use size_t instead of uint
- Functions that used a pointer to a string length was changed to use size_t*
- Changed malloc(), free() and related functions from using gptr to use void *
as this requires fewer casts in the code and is more in line with how the
standard functions work.
- Added extra length argument to dirname_part() to return the length of the
created string.
- Changed (at least) following functions to take uchar* as argument:
- db_dump()
- my_net_write()
- net_write_command()
- net_store_data()
- DBUG_DUMP()
- decimal2bin() & bin2decimal()
- Changed my_compress() and my_uncompress() to use size_t. Changed one
argument to my_uncompress() from a pointer to a value as we only return
one value (makes function easier to use).
- Changed type of 'pack_data' argument to packfrm() to avoid casts.
- Changed in readfrm() and writefrom(), ha_discover and handler::discover()
the type for argument 'frmdata' to uchar** to avoid casts.
- Changed most Field functions to use uchar* instead of char* (reduced a lot of
casts).
- Changed field->val_xxx(xxx, new_ptr) to take const pointers.
Other changes:
- Removed a lot of not needed casts
- Added a few new cast required by other changes
- Added some cast to my_multi_malloc() arguments for safety (as string lengths
needs to be uint, not size_t).
- Fixed all calls to hash-get-key functions to use size_t*. (Needed to be done
explicitely as this conflict was often hided by casting the function to
hash_get_key).
- Changed some buffers to memory regions to uchar* to avoid casts.
- Changed some string lengths from uint to size_t.
- Changed field->ptr to be uchar* instead of char*. This allowed us to
get rid of a lot of casts.
- Some changes from true -> TRUE, false -> FALSE, unsigned char -> uchar
- Include zlib.h in some files as we needed declaration of crc32()
- Changed MY_FILE_ERROR to be (size_t) -1.
- Changed many variables to hold the result of my_read() / my_write() to be
size_t. This was needed to properly detect errors (which are
returned as (size_t) -1).
- Removed some very old VMS code
- Changed packfrm()/unpackfrm() to not be depending on uint size
(portability fix)
- Removed windows specific code to restore cursor position as this
causes slowdown on windows and we should not mix read() and pread()
calls anyway as this is not thread safe. Updated function comment to
reflect this. Changed function that depended on original behavior of
my_pwrite() to itself restore the cursor position (one such case).
- Added some missing checking of return value of malloc().
- Changed definition of MOD_PAD_CHAR_TO_FULL_LENGTH to avoid 'long' overflow.
- Changed type of table_def::m_size from my_size_t to ulong to reflect that
m_size is the number of elements in the array, not a string/memory
length.
- Moved THD::max_row_length() to table.cc (as it's not depending on THD).
Inlined max_row_length_blob() into this function.
- More function comments
- Fixed some compiler warnings when compiled without partitions.
- Removed setting of LEX_STRING() arguments in declaration (portability fix).
- Some trivial indentation/variable name changes.
- Some trivial code simplifications:
- Replaced some calls to alloc_root + memcpy to use
strmake_root()/strdup_root().
- Changed some calls from memdup() to strmake() (Safety fix)
- Simpler loops in client-simple.c
2007-05-10 11:59:39 +02:00
|
|
|
if (!bcmp((uchar*) reg_ext, (uchar*) ext, ext_len))
|
2007-01-10 09:57:03 +01:00
|
|
|
{
|
2007-01-12 10:34:33 +01:00
|
|
|
handler *handler_file= 0;
|
|
|
|
/* We should cut file extention before deleting of table */
|
|
|
|
memcpy(filePathCopy, filePath, filePath_len - ext_len);
|
|
|
|
filePathCopy[filePath_len - ext_len]= 0;
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
init_tmp_table_share(thd, &share, "", 0, "", filePathCopy);
|
2007-01-12 10:34:33 +01:00
|
|
|
if (!open_table_def(thd, &share, 0) &&
|
|
|
|
((handler_file= get_new_handler(&share, thd->mem_root,
|
2007-03-02 17:43:45 +01:00
|
|
|
share.db_type()))))
|
2007-01-12 10:34:33 +01:00
|
|
|
{
|
2007-12-20 19:16:55 +01:00
|
|
|
handler_file->ha_delete_table(filePathCopy);
|
2007-01-12 10:34:33 +01:00
|
|
|
delete handler_file;
|
|
|
|
}
|
|
|
|
free_table_share(&share);
|
2007-01-10 09:57:03 +01:00
|
|
|
}
|
2007-01-12 10:34:33 +01:00
|
|
|
/*
|
|
|
|
File can be already deleted by tmp_table.file->delete_table().
|
|
|
|
So we hide error messages which happnes during deleting of these
|
|
|
|
files(MYF(0)).
|
|
|
|
*/
|
2009-11-24 14:54:59 +01:00
|
|
|
(void) my_delete(filePath, MYF(0));
|
2005-11-23 21:45:02 +01:00
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
2005-11-23 21:45:02 +01:00
|
|
|
my_dirend(dirp);
|
2002-10-08 14:39:37 +02:00
|
|
|
}
|
2007-01-10 09:57:03 +01:00
|
|
|
delete thd;
|
|
|
|
my_pthread_setspecific_ptr(THR_THD, 0);
|
|
|
|
DBUG_RETURN(0);
|
2000-07-31 21:29:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/*****************************************************************************
|
|
|
|
unireg support functions
|
|
|
|
*****************************************************************************/
|
|
|
|
|
|
|
|
/*
|
2005-11-23 21:45:02 +01:00
|
|
|
free all unused tables
|
|
|
|
|
|
|
|
NOTE
|
|
|
|
This is called by 'handle_manager' when one wants to periodicly flush
|
|
|
|
all not used tables.
|
2000-07-31 21:29:14 +02:00
|
|
|
*/
|
|
|
|
|
|
|
|
void flush_tables()
|
|
|
|
{
|
|
|
|
(void) pthread_mutex_lock(&LOCK_open);
|
|
|
|
while (unused_tables)
|
2009-11-30 16:55:03 +01:00
|
|
|
free_cache_entry(unused_tables);
|
2000-07-31 21:29:14 +02:00
|
|
|
(void) pthread_mutex_unlock(&LOCK_open);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
/**
|
|
|
|
A callback to the server internals that is used to address
|
|
|
|
special cases of the locking protocol.
|
|
|
|
Invoked when acquiring an exclusive lock, for each thread that
|
|
|
|
has a conflicting shared metadata lock.
|
|
|
|
|
|
|
|
This function:
|
|
|
|
- aborts waiting of the thread on a data lock, to make it notice
|
|
|
|
the pending exclusive lock and back off.
|
|
|
|
- if the thread is an INSERT DELAYED thread, sends it a KILL
|
|
|
|
signal to terminate it.
|
|
|
|
|
|
|
|
@note This function does not wait for the thread to give away its
|
|
|
|
locks. Waiting is done outside for all threads at once.
|
|
|
|
|
|
|
|
@param thd Current thread context
|
|
|
|
@param in_use The thread to wake up
|
|
|
|
|
|
|
|
@retval TRUE if the thread was woken up
|
|
|
|
@retval FALSE otherwise (e.g. it was not waiting for a table-level lock).
|
|
|
|
|
|
|
|
@note It is one of two places where border between MDL and the
|
|
|
|
rest of the server is broken.
|
|
|
|
*/
|
|
|
|
|
2009-12-02 17:31:57 +01:00
|
|
|
bool mysql_notify_thread_having_shared_lock(THD *thd, THD *in_use)
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
|
|
|
bool signalled= FALSE;
|
|
|
|
if ((in_use->system_thread & SYSTEM_THREAD_DELAYED_INSERT) &&
|
|
|
|
!in_use->killed)
|
|
|
|
{
|
|
|
|
in_use->killed= THD::KILL_CONNECTION;
|
|
|
|
pthread_mutex_lock(&in_use->mysys_var->mutex);
|
|
|
|
if (in_use->mysys_var->current_cond)
|
|
|
|
pthread_cond_broadcast(in_use->mysys_var->current_cond);
|
|
|
|
pthread_mutex_unlock(&in_use->mysys_var->mutex);
|
|
|
|
signalled= TRUE;
|
|
|
|
}
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
for (TABLE *thd_table= in_use->open_tables;
|
|
|
|
thd_table ;
|
|
|
|
thd_table= thd_table->next)
|
|
|
|
{
|
2009-12-01 16:20:43 +01:00
|
|
|
/*
|
|
|
|
Check for TABLE::db_stat is needed since in some places we call
|
|
|
|
handler::close() for table instance (and set TABLE::db_stat to 0)
|
|
|
|
and do not remove such instances from the THD::open_tables
|
|
|
|
for some time, during which other thread can see those instances
|
|
|
|
(e.g. see partitioning code).
|
|
|
|
*/
|
2009-11-30 16:55:03 +01:00
|
|
|
if (thd_table->db_stat)
|
|
|
|
signalled|= mysql_lock_abort_for_thread(thd, thd_table);
|
|
|
|
}
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
return signalled;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
2009-12-01 20:13:01 +01:00
|
|
|
Remove all or some (depending on parameter) instances of TABLE and
|
|
|
|
TABLE_SHARE from the table definition cache.
|
|
|
|
|
|
|
|
@param thd Thread context
|
|
|
|
@param remove_type Type of removal:
|
|
|
|
TDC_RT_REMOVE_ALL - remove all TABLE instances and
|
|
|
|
TABLE_SHARE instance. There
|
|
|
|
should be no used TABLE objects
|
|
|
|
and caller should have exclusive
|
|
|
|
metadata lock on the table.
|
|
|
|
TDC_RT_REMOVE_NOT_OWN - remove all TABLE instances
|
|
|
|
except those that belong to
|
|
|
|
this thread. There should be
|
|
|
|
no TABLE objects used by other
|
|
|
|
threads and caller should have
|
|
|
|
exclusive metadata lock on the
|
|
|
|
table.
|
|
|
|
TDC_RT_REMOVE_UNUSED - remove all unused TABLE
|
|
|
|
instances (if there are no
|
|
|
|
used instances will also
|
|
|
|
remove TABLE_SHARE).
|
|
|
|
@param db Name of database
|
|
|
|
@param table_name Name of table
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
@note Unlike remove_table_from_cache() it assumes that table instances
|
|
|
|
are already not used by any (other) thread (this should be achieved
|
|
|
|
by using meta-data locks).
|
|
|
|
*/
|
|
|
|
|
2009-12-01 20:13:01 +01:00
|
|
|
void tdc_remove_table(THD *thd, enum_tdc_remove_table_type remove_type,
|
|
|
|
const char *db, const char *table_name)
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
|
|
|
char key[MAX_DBKEY_LENGTH];
|
|
|
|
uint key_length;
|
|
|
|
TABLE *table;
|
|
|
|
TABLE_SHARE *share;
|
|
|
|
|
2009-12-01 20:13:01 +01:00
|
|
|
safe_mutex_assert_owner(&LOCK_open);
|
2009-11-30 16:55:03 +01:00
|
|
|
|
2009-12-01 20:13:01 +01:00
|
|
|
DBUG_ASSERT(remove_type == TDC_RT_REMOVE_UNUSED ||
|
2009-12-10 09:21:38 +01:00
|
|
|
thd->mdl_context.is_exclusive_lock_owner(MDL_key::TABLE,
|
|
|
|
db, table_name));
|
2009-11-30 16:55:03 +01:00
|
|
|
|
2009-12-01 20:13:01 +01:00
|
|
|
key_length=(uint) (strmov(strmov(key,db)+1,table_name)-key)+1;
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
if ((share= (TABLE_SHARE*) my_hash_search(&table_def_cache,(uchar*) key,
|
|
|
|
key_length)))
|
|
|
|
{
|
2009-12-01 20:13:01 +01:00
|
|
|
if (share->ref_count)
|
|
|
|
{
|
|
|
|
I_P_List_iterator<TABLE, TABLE_share> it(share->free_tables);
|
|
|
|
#ifndef DBUG_OFF
|
|
|
|
if (remove_type == TDC_RT_REMOVE_ALL)
|
|
|
|
{
|
|
|
|
DBUG_ASSERT(share->used_tables.is_empty());
|
|
|
|
}
|
|
|
|
else if (remove_type == TDC_RT_REMOVE_NOT_OWN)
|
|
|
|
{
|
|
|
|
I_P_List_iterator<TABLE, TABLE_share> it2(share->used_tables);
|
|
|
|
while ((table= it2++))
|
|
|
|
if (table->in_use != thd)
|
|
|
|
{
|
|
|
|
DBUG_ASSERT(0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
/*
|
|
|
|
Set share's version to zero in order to ensure that it gets
|
|
|
|
automatically deleted once it is no longer referenced.
|
|
|
|
*/
|
|
|
|
share->version= 0;
|
|
|
|
while ((table= it++))
|
|
|
|
free_cache_entry(table);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
(void) my_hash_delete(&table_def_cache, (uchar*) share);
|
2009-11-30 16:55:03 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
Wait until there are no old versions of tables in the table
|
|
|
|
definition cache for the metadata locks that we try to acquire.
|
|
|
|
|
|
|
|
@param thd Thread context
|
|
|
|
@param context Metadata locking context with locks.
|
|
|
|
*/
|
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
static bool
|
|
|
|
tdc_wait_for_old_versions(THD *thd, MDL_request_list *mdl_requests)
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
|
|
|
TABLE_SHARE *share;
|
|
|
|
const char *old_msg;
|
2009-12-04 00:52:05 +01:00
|
|
|
MDL_request *mdl_request;
|
2009-11-30 16:55:03 +01:00
|
|
|
|
|
|
|
while (!thd->killed)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
Here we have situation as in mdl_wait_for_locks() we need to
|
|
|
|
get rid of offending HANDLERs to avoid deadlock.
|
|
|
|
TODO: We should also investigate in which situations we have
|
|
|
|
to broadcast on COND_refresh because of this.
|
|
|
|
*/
|
|
|
|
mysql_ha_flush(thd);
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
|
2009-12-08 10:57:07 +01:00
|
|
|
MDL_request_list::Iterator it(*mdl_requests);
|
2009-12-04 00:52:05 +01:00
|
|
|
while ((mdl_request= it++))
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
2009-12-04 00:52:05 +01:00
|
|
|
if ((share= get_cached_table_share(mdl_request->key.db_name(),
|
2009-12-09 17:11:26 +01:00
|
|
|
mdl_request->key.name())) &&
|
2009-11-30 16:55:03 +01:00
|
|
|
share->version != refresh_version &&
|
|
|
|
!share->used_tables.is_empty())
|
|
|
|
break;
|
|
|
|
}
|
2009-12-04 00:52:05 +01:00
|
|
|
if (!mdl_request)
|
2009-11-30 16:55:03 +01:00
|
|
|
{
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
old_msg= thd->enter_cond(&COND_refresh, &LOCK_open, "Waiting for table");
|
|
|
|
pthread_cond_wait(&COND_refresh, &LOCK_open);
|
|
|
|
/* LOCK_open mutex is unlocked by THD::exit_cond() as side-effect. */
|
|
|
|
thd->exit_cond(old_msg);
|
|
|
|
}
|
|
|
|
return thd->killed;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2002-10-30 12:18:52 +01:00
|
|
|
int setup_ftfuncs(SELECT_LEX *select_lex)
|
2000-07-31 21:29:14 +02:00
|
|
|
{
|
2002-10-30 12:18:52 +01:00
|
|
|
List_iterator<Item_func_match> li(*(select_lex->ftfunc_list)),
|
|
|
|
lj(*(select_lex->ftfunc_list));
|
2000-08-28 15:43:58 +02:00
|
|
|
Item_func_match *ftf, *ftf2;
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
while ((ftf=li++))
|
2000-08-28 15:43:58 +02:00
|
|
|
{
|
2000-07-31 21:29:14 +02:00
|
|
|
if (ftf->fix_index())
|
|
|
|
return 1;
|
2001-11-21 18:05:11 +01:00
|
|
|
lj.rewind();
|
|
|
|
while ((ftf2=lj++) != ftf)
|
2000-08-28 15:43:58 +02:00
|
|
|
{
|
2002-03-22 13:03:42 +01:00
|
|
|
if (ftf->eq(ftf2,1) && !ftf2->master)
|
2000-08-28 15:43:58 +02:00
|
|
|
ftf2->master=ftf;
|
|
|
|
}
|
|
|
|
}
|
2000-07-31 21:29:14 +02:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
2001-11-21 18:05:11 +01:00
|
|
|
|
2002-08-05 17:10:53 +02:00
|
|
|
|
2002-10-30 12:18:52 +01:00
|
|
|
int init_ftfuncs(THD *thd, SELECT_LEX *select_lex, bool no_order)
|
2001-11-21 18:05:11 +01:00
|
|
|
{
|
2002-10-30 12:18:52 +01:00
|
|
|
if (select_lex->ftfunc_list->elements)
|
2001-11-21 18:05:11 +01:00
|
|
|
{
|
2002-10-30 12:18:52 +01:00
|
|
|
List_iterator<Item_func_match> li(*(select_lex->ftfunc_list));
|
2001-12-06 00:05:30 +01:00
|
|
|
Item_func_match *ifm;
|
|
|
|
DBUG_PRINT("info",("Performing FULLTEXT search"));
|
2007-02-22 16:03:08 +01:00
|
|
|
thd_proc_info(thd, "FULLTEXT initialization");
|
2001-11-21 18:05:11 +01:00
|
|
|
|
2001-12-06 00:05:30 +01:00
|
|
|
while ((ifm=li++))
|
|
|
|
ifm->init_search(no_order);
|
|
|
|
}
|
2001-11-21 18:05:11 +01:00
|
|
|
return 0;
|
|
|
|
}
|
2004-07-16 00:15:55 +02:00
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
open new .frm format table
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
open_new_frm()
|
2005-11-01 14:54:30 +01:00
|
|
|
THD thread handler
|
2005-11-23 21:45:02 +01:00
|
|
|
path path to .frm file (without extension)
|
2004-07-16 00:15:55 +02:00
|
|
|
alias alias for table
|
2004-08-24 14:37:51 +02:00
|
|
|
db database
|
|
|
|
table_name name of table
|
This changeset is largely a handler cleanup changeset (WL#3281), but includes fixes and cleanups that was found necessary while testing the handler changes
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
2006-06-04 17:52:22 +02:00
|
|
|
db_stat open flags (for example ->OPEN_KEYFILE|HA_OPEN_RNDFILE..)
|
2004-07-16 00:15:55 +02:00
|
|
|
can be 0 (example in ha_example_table)
|
|
|
|
prgflag READ_ALL etc..
|
|
|
|
ha_open_flags HA_OPEN_ABORT_IF_LOCKED etc..
|
|
|
|
outparam result table
|
|
|
|
table_desc TABLE_LIST descriptor
|
|
|
|
mem_root temporary MEM_ROOT for parsing
|
|
|
|
*/
|
2004-09-03 20:43:04 +02:00
|
|
|
|
2005-11-01 14:54:30 +01:00
|
|
|
static bool
|
2005-11-23 21:45:02 +01:00
|
|
|
open_new_frm(THD *thd, TABLE_SHARE *share, const char *alias,
|
2004-08-24 14:37:51 +02:00
|
|
|
uint db_stat, uint prgflag,
|
2004-07-16 00:15:55 +02:00
|
|
|
uint ha_open_flags, TABLE *outparam, TABLE_LIST *table_desc,
|
|
|
|
MEM_ROOT *mem_root)
|
|
|
|
{
|
|
|
|
LEX_STRING pathstr;
|
2004-09-03 20:43:04 +02:00
|
|
|
File_parser *parser;
|
2005-11-23 21:45:02 +01:00
|
|
|
char path[FN_REFLEN];
|
2004-09-03 20:43:04 +02:00
|
|
|
DBUG_ENTER("open_new_frm");
|
|
|
|
|
2005-11-23 21:45:02 +01:00
|
|
|
/* Create path with extension */
|
|
|
|
pathstr.length= (uint) (strxmov(path, share->normalized_path.str, reg_ext,
|
|
|
|
NullS)- path);
|
|
|
|
pathstr.str= path;
|
2004-08-24 14:37:51 +02:00
|
|
|
|
2004-09-03 20:43:04 +02:00
|
|
|
if ((parser= sql_parse_prepare(&pathstr, mem_root, 1)))
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2005-07-31 11:49:55 +02:00
|
|
|
if (is_equal(&view_type, parser->type()))
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2004-08-24 14:37:51 +02:00
|
|
|
if (table_desc == 0 || table_desc->required_type == FRMTYPE_TABLE)
|
2004-07-16 00:15:55 +02:00
|
|
|
{
|
2005-11-23 21:45:02 +01:00
|
|
|
my_error(ER_WRONG_OBJECT, MYF(0), share->db.str, share->table_name.str,
|
|
|
|
"BASE TABLE");
|
2004-08-24 14:37:51 +02:00
|
|
|
goto err;
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
2006-08-08 22:05:42 +02:00
|
|
|
if (mysql_make_view(thd, parser, table_desc,
|
|
|
|
(prgflag & OPEN_VIEW_NO_PARSE)))
|
2004-08-24 14:37:51 +02:00
|
|
|
goto err;
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
/* only VIEWs are supported now */
|
2007-10-25 03:01:08 +02:00
|
|
|
my_error(ER_FRM_UNKNOWN_TYPE, MYF(0), share->path.str, parser->type()->str);
|
2004-08-24 14:37:51 +02:00
|
|
|
goto err;
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
2004-09-03 20:43:04 +02:00
|
|
|
DBUG_RETURN(0);
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
2004-09-03 21:38:45 +02:00
|
|
|
|
2004-08-24 14:37:51 +02:00
|
|
|
err:
|
|
|
|
DBUG_RETURN(1);
|
2004-07-16 00:15:55 +02:00
|
|
|
}
|
2005-07-31 11:49:55 +02:00
|
|
|
|
|
|
|
|
|
|
|
bool is_equal(const LEX_STRING *a, const LEX_STRING *b)
|
|
|
|
{
|
|
|
|
return a->length == b->length && !strncmp(a->str, b->str, a->length);
|
|
|
|
}
|
2006-01-17 08:40:00 +01:00
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
SYNOPSIS
|
|
|
|
abort_and_upgrade_lock()
|
|
|
|
lpt Parameter passing struct
|
|
|
|
All parameters passed through the ALTER_PARTITION_PARAM_TYPE object
|
2006-04-16 03:49:13 +02:00
|
|
|
RETURN VALUE
|
|
|
|
0
|
2006-01-17 08:40:00 +01:00
|
|
|
DESCRIPTION
|
|
|
|
Remember old lock level (for possible downgrade later on), abort all
|
|
|
|
waiting threads and ensure that all keeping locks currently are
|
|
|
|
completed such that we own the lock exclusively and no other interaction
|
|
|
|
is ongoing.
|
|
|
|
|
|
|
|
thd Thread object
|
|
|
|
table Table object
|
|
|
|
db Database name
|
|
|
|
table_name Table name
|
|
|
|
old_lock_level Old lock level
|
|
|
|
*/
|
|
|
|
|
2006-04-16 03:49:13 +02:00
|
|
|
int abort_and_upgrade_lock(ALTER_PARTITION_PARAM_TYPE *lpt)
|
2006-01-17 08:40:00 +01:00
|
|
|
{
|
2009-12-02 16:22:15 +01:00
|
|
|
DBUG_ENTER("abort_and_upgrade_lock");
|
|
|
|
|
|
|
|
if (wait_while_table_is_used(lpt->thd, lpt->table, HA_EXTRA_FORCE_REOPEN))
|
2009-11-30 16:55:03 +01:00
|
|
|
DBUG_RETURN(1);
|
2006-04-16 03:49:13 +02:00
|
|
|
DBUG_RETURN(0);
|
2006-01-17 08:40:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
2009-11-30 16:55:03 +01:00
|
|
|
Tells if two (or more) tables have auto_increment columns and we want to
|
|
|
|
lock those tables with a write lock.
|
2009-10-01 01:19:36 +02:00
|
|
|
|
2009-11-30 16:55:03 +01:00
|
|
|
SYNOPSIS
|
|
|
|
has_two_write_locked_tables_with_auto_increment
|
|
|
|
tables Table list
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
|
|
|
|
NOTES:
|
|
|
|
Call this function only when you have established the list of all tables
|
|
|
|
which you'll want to update (including stored functions, triggers, views
|
|
|
|
inside your statement).
|
|
|
|
*/
|
|
|
|
|
|
|
|
static bool
|
2009-10-01 01:19:36 +02:00
|
|
|
has_write_table_with_auto_increment(TABLE_LIST *tables)
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
{
|
|
|
|
for (TABLE_LIST *table= tables; table; table= table->next_global)
|
|
|
|
{
|
|
|
|
/* we must do preliminary checks as table->table may be NULL */
|
2007-05-11 19:51:03 +02:00
|
|
|
if (!table->placeholder() &&
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
table->table->found_next_number_field &&
|
|
|
|
(table->lock_type >= TL_WRITE_ALLOW_WRITE))
|
2009-10-01 01:19:36 +02:00
|
|
|
return 1;
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
}
|
2009-10-01 01:19:36 +02:00
|
|
|
|
* Mixed replication mode * :
1) Fix for BUG#19630 "stored function inserting into two auto_increment breaks
statement-based binlog":
a stored function inserting into two such tables may fail to replicate
(inserting wrong data in the slave's copy of the second table) if the slave's
second table had an internal auto_increment counter different from master's.
Because the auto_increment value autogenerated by master for the 2nd table
does not go into binlog, only the first does, so the slave lacks information.
To fix this, if running in mixed binlogging mode, if the stored function or
trigger plans to update two different tables both having auto_increment
columns, we switch to row-based for the whole function.
We don't have a simple solution for statement-based binlogging mode, there
the bug remains and will be documented as a known problem.
Re-enabling rpl_switch_stm_row_mixed.
2) Fix for BUG#20630 "Mixed binlogging mode does not work with stored
functions, triggers, views", which was a documented limitation (in mixed
mode, we didn't detect that a stored function's execution needed row-based
binlogging (due to some UUID() call for example); same for
triggers, same for views (a view created from a SELECT UUID(), and doing
INSERT INTO sometable SELECT theview; would not replicate row-based).
This is implemented by, after parsing a routine's body, remembering in sp_head
that this routine needs row-based binlogging. Then when this routine is used,
the caller is marked to require row-based binlogging too.
Same for views: when we parse a view and detect that its SELECT needs
row-based binary logging, we mark the calling LEX as such.
3) Fix for BUG#20499 "mixed mode with temporary table breaks binlog":
a temporary table containing e.g. UUID has its changes not binlogged,
so any query updating a permanent table with data from the temporary table
will run wrongly on slave. Solution: in mixed mode we don't switch back
from row-based to statement-based when there exists temporary tables.
4) Attempt to test mysqlbinlog on a binlog generated by mysqlbinlog;
impossible due to BUG#11312 and BUG#20329, but test is in place for when
they are fixed.
2006-07-09 17:00:47 +02:00
|
|
|
return 0;
|
|
|
|
}
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Open and lock system tables for read.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
open_system_tables_for_read()
|
|
|
|
thd Thread context.
|
|
|
|
table_list List of tables to open.
|
|
|
|
backup Pointer to Open_tables_state instance where
|
|
|
|
information about currently open tables will be
|
|
|
|
saved, and from which will be restored when we will
|
|
|
|
end work with system tables.
|
|
|
|
|
|
|
|
NOTES
|
|
|
|
Thanks to restrictions which we put on opening and locking of
|
|
|
|
system tables for writing, we can open and lock them for reading
|
|
|
|
even when we already have some other tables open and locked. One
|
|
|
|
must call close_system_tables() to close systems tables opened
|
|
|
|
with this call.
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
FALSE Success
|
|
|
|
TRUE Error
|
|
|
|
*/
|
|
|
|
|
|
|
|
bool
|
|
|
|
open_system_tables_for_read(THD *thd, TABLE_LIST *table_list,
|
|
|
|
Open_tables_state *backup)
|
|
|
|
{
|
2009-11-30 20:03:37 +01:00
|
|
|
Query_tables_list query_tables_list_backup;
|
|
|
|
LEX *lex= thd->lex;
|
|
|
|
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
DBUG_ENTER("open_system_tables_for_read");
|
|
|
|
|
2009-11-30 20:03:37 +01:00
|
|
|
/*
|
|
|
|
Besides using new Open_tables_state for opening system tables,
|
|
|
|
we also have to backup and reset/and then restore part of LEX
|
|
|
|
which is accessed by open_tables() in order to determine if
|
|
|
|
prelocking is needed and what tables should be added for it.
|
|
|
|
close_system_tables() doesn't require such treatment.
|
|
|
|
*/
|
|
|
|
lex->reset_n_backup_query_tables_list(&query_tables_list_backup);
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
thd->reset_n_backup_open_tables_state(backup);
|
|
|
|
|
2009-11-30 20:03:37 +01:00
|
|
|
if (open_and_lock_tables_derived(thd, table_list, FALSE,
|
|
|
|
MYSQL_LOCK_IGNORE_FLUSH))
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
{
|
2009-11-30 20:03:37 +01:00
|
|
|
lex->restore_backup_query_tables_list(&query_tables_list_backup);
|
|
|
|
goto error;
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
}
|
|
|
|
|
2009-11-30 20:03:37 +01:00
|
|
|
for (TABLE_LIST *tables= table_list; tables; tables= tables->next_global)
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
{
|
2009-11-30 20:03:37 +01:00
|
|
|
DBUG_ASSERT(tables->table->s->table_category == TABLE_CATEGORY_SYSTEM);
|
|
|
|
tables->table->use_all_columns();
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
}
|
2009-11-30 20:03:37 +01:00
|
|
|
lex->restore_backup_query_tables_list(&query_tables_list_backup);
|
|
|
|
|
|
|
|
DBUG_RETURN(FALSE);
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
|
|
|
|
error:
|
|
|
|
close_system_tables(thd, backup);
|
|
|
|
|
|
|
|
DBUG_RETURN(TRUE);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Close system tables, opened with open_system_tables_for_read().
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
close_system_tables()
|
|
|
|
thd Thread context
|
|
|
|
backup Pointer to Open_tables_state instance which holds
|
|
|
|
information about tables which were open before we
|
|
|
|
decided to access system tables.
|
|
|
|
*/
|
|
|
|
|
|
|
|
void
|
|
|
|
close_system_tables(THD *thd, Open_tables_state *backup)
|
|
|
|
{
|
|
|
|
close_thread_tables(thd);
|
|
|
|
thd->restore_backup_open_tables_state(backup);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
Open and lock one system table for update.
|
|
|
|
|
|
|
|
SYNOPSIS
|
|
|
|
open_system_table_for_update()
|
|
|
|
thd Thread context.
|
|
|
|
one_table Table to open.
|
|
|
|
|
|
|
|
NOTES
|
|
|
|
Table opened with this call should closed using close_thread_tables().
|
|
|
|
|
|
|
|
RETURN
|
|
|
|
0 Error
|
|
|
|
# Pointer to TABLE object of system table
|
|
|
|
*/
|
|
|
|
|
|
|
|
TABLE *
|
|
|
|
open_system_table_for_update(THD *thd, TABLE_LIST *one_table)
|
|
|
|
{
|
|
|
|
DBUG_ENTER("open_system_table_for_update");
|
|
|
|
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
TABLE *table= open_ltable(thd, one_table, one_table->lock_type, 0);
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
if (table)
|
|
|
|
{
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
DBUG_ASSERT(table->s->table_category == TABLE_CATEGORY_SYSTEM);
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
table->use_all_columns();
|
|
|
|
}
|
|
|
|
|
|
|
|
DBUG_RETURN(table);
|
|
|
|
}
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
|
|
|
|
/**
|
|
|
|
Open a performance schema table.
|
|
|
|
Opening such tables is performed internally in the server
|
|
|
|
implementation, and is a 'nested' open, since some tables
|
|
|
|
might be already opened by the current thread.
|
|
|
|
The thread context before this call is saved, and is restored
|
|
|
|
when calling close_performance_schema_table().
|
|
|
|
@param thd The current thread
|
|
|
|
@param one_table Performance schema table to open
|
|
|
|
@param backup [out] Temporary storage used to save the thread context
|
|
|
|
*/
|
|
|
|
TABLE *
|
|
|
|
open_performance_schema_table(THD *thd, TABLE_LIST *one_table,
|
|
|
|
Open_tables_state *backup)
|
|
|
|
{
|
2007-08-03 00:14:27 +02:00
|
|
|
uint flags= ( MYSQL_LOCK_IGNORE_GLOBAL_READ_LOCK |
|
|
|
|
MYSQL_LOCK_IGNORE_GLOBAL_READ_ONLY |
|
|
|
|
MYSQL_LOCK_IGNORE_FLUSH |
|
|
|
|
MYSQL_LOCK_PERF_SCHEMA);
|
|
|
|
TABLE *table;
|
|
|
|
/* Save value that is changed in mysql_lock_tables() */
|
|
|
|
ulonglong save_utime_after_lock= thd->utime_after_lock;
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
DBUG_ENTER("open_performance_schema_table");
|
|
|
|
|
|
|
|
thd->reset_n_backup_open_tables_state(backup);
|
|
|
|
|
2007-08-03 00:14:27 +02:00
|
|
|
if ((table= open_ltable(thd, one_table, one_table->lock_type, flags)))
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
{
|
|
|
|
DBUG_ASSERT(table->s->table_category == TABLE_CATEGORY_PERFORMANCE);
|
|
|
|
/* Make sure all columns get assigned to a default value */
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
table->use_all_columns();
|
2007-07-30 19:56:02 +02:00
|
|
|
table->no_replicate= 1;
|
2007-08-03 00:14:27 +02:00
|
|
|
/*
|
|
|
|
Don't set automatic timestamps as we may want to use time of logging,
|
|
|
|
not from query start
|
|
|
|
*/
|
|
|
|
table->timestamp_field_type= TIMESTAMP_NO_AUTO_SET;
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
}
|
2007-08-30 02:56:00 +02:00
|
|
|
else
|
2007-10-18 16:14:27 +02:00
|
|
|
{
|
|
|
|
/*
|
|
|
|
If error in mysql_lock_tables(), open_ltable doesn't close the
|
|
|
|
table. Thread kill during mysql_lock_tables() is such error. But
|
|
|
|
open tables cannot be accepted when restoring the open tables
|
|
|
|
state.
|
|
|
|
*/
|
|
|
|
if (thd->killed)
|
|
|
|
close_thread_tables(thd);
|
2007-08-30 02:56:00 +02:00
|
|
|
thd->restore_backup_open_tables_state(backup);
|
2007-10-18 16:14:27 +02:00
|
|
|
}
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
|
2007-08-03 00:14:27 +02:00
|
|
|
thd->utime_after_lock= save_utime_after_lock;
|
BUG#9953: CONVERT_TZ requires mysql.time_zone_name to be locked
The problem was that some facilities (like CONVERT_TZ() function or
server HELP statement) may require implicit access to some tables in
'mysql' database. This access was done by ordinary means of adding
such tables to the list of tables the query is going to open.
However, if we issued LOCK TABLES before that, we would get "table
was not locked" error trying to open such implicit tables.
The solution is to treat certain tables as MySQL system tables, like
we already do for mysql.proc. Such tables may be opened for reading
at any moment regardless of any locks in effect. The cost of this is
that system table may be locked for writing only together with other
system tables, it is disallowed to lock system tables for writing and
have any other lock on any other table.
After this patch the following tables are treated as MySQL system
tables:
mysql.help_category
mysql.help_keyword
mysql.help_relation
mysql.help_topic
mysql.proc (it already was)
mysql.time_zone
mysql.time_zone_leap_second
mysql.time_zone_name
mysql.time_zone_transition
mysql.time_zone_transition_type
These tables are now opened with open_system_tables_for_read() and
closed with close_system_tables(), or one table may be opened with
open_system_table_for_update() and closed with close_thread_tables()
(the latter is used for mysql.proc table, which is updated as part of
normal MySQL server operation). These functions may be used when
some tables were opened and locked already.
NOTE: online update of time zone tables is not possible during
replication, because there's no time zone cache flush neither on LOCK
TABLES, nor on FLUSH TABLES, so the master may serve stale time zone
data from cache, while on slave updated data will be loaded from the
time zone tables.
2007-03-09 11:12:31 +01:00
|
|
|
DBUG_RETURN(table);
|
|
|
|
}
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
|
|
|
|
/**
|
|
|
|
Close a performance schema table.
|
|
|
|
The last table opened by open_performance_schema_table()
|
|
|
|
is closed, then the thread context is restored.
|
|
|
|
@param thd The current thread
|
|
|
|
@param backup [in] the context to restore.
|
|
|
|
*/
|
|
|
|
void close_performance_schema_table(THD *thd, Open_tables_state *backup)
|
|
|
|
{
|
|
|
|
bool found_old_table;
|
|
|
|
|
2007-08-30 02:56:00 +02:00
|
|
|
/*
|
|
|
|
If open_performance_schema_table() fails,
|
|
|
|
this function should not be called.
|
|
|
|
*/
|
|
|
|
DBUG_ASSERT(thd->lock != NULL);
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
|
2007-08-30 02:56:00 +02:00
|
|
|
/*
|
|
|
|
Note:
|
|
|
|
We do not create explicitly a separate transaction for the
|
|
|
|
performance table I/O, but borrow the current transaction.
|
|
|
|
lock + unlock will autocommit the change done in the
|
|
|
|
performance schema table: this is the expected result.
|
|
|
|
The current transaction should not be affected by this code.
|
|
|
|
TODO: Note that if a transactional engine is used for log tables,
|
|
|
|
this code will need to be revised, as a separate transaction
|
|
|
|
might be needed.
|
|
|
|
*/
|
|
|
|
mysql_unlock_tables(thd, thd->lock);
|
|
|
|
thd->lock= 0;
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
|
|
|
|
pthread_mutex_lock(&LOCK_open);
|
|
|
|
|
|
|
|
found_old_table= false;
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
/*
|
|
|
|
Note that we need to hold LOCK_open while changing the
|
|
|
|
open_tables list. Another thread may work on it.
|
2009-12-03 00:09:22 +01:00
|
|
|
(See: notify_thread_having_shared_lock())
|
Bug#26379 - Combination of FLUSH TABLE and REPAIR TABLE
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.
2007-11-15 20:25:43 +01:00
|
|
|
*/
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
while (thd->open_tables)
|
|
|
|
found_old_table|= close_thread_table(thd, &thd->open_tables);
|
|
|
|
|
|
|
|
if (found_old_table)
|
|
|
|
broadcast_refresh();
|
|
|
|
|
|
|
|
pthread_mutex_unlock(&LOCK_open);
|
|
|
|
|
2009-12-04 00:52:05 +01:00
|
|
|
thd->mdl_context.release_all_locks();
|
2009-11-30 16:55:03 +01:00
|
|
|
|
WL#3984 (Revise locking of mysql.general_log and mysql.slow_log)
Bug#25422 (Hang with log tables)
Bug 17876 (Truncating mysql.slow_log in a SP after using cursor locks the
thread)
Bug 23044 (Warnings on flush of a log table)
Bug 29129 (Resetting general_log while the GLOBAL READ LOCK is set causes
a deadlock)
Prior to this fix, the server would hang when performing concurrent
ALTER TABLE or TRUNCATE TABLE statements against the LOG TABLES,
which are mysql.general_log and mysql.slow_log.
The root cause traces to the following code:
in sql_base.cc, open_table()
if (table->in_use != thd)
{
/* wait_for_condition will unlock LOCK_open for us */
wait_for_condition(thd, &LOCK_open, &COND_refresh);
}
The problem with this code is that the current implementation of the
LOGGER creates 'fake' THD objects, like
- Log_to_csv_event_handler::general_log_thd
- Log_to_csv_event_handler::slow_log_thd
which are not associated to a real thread running in the server,
so that waiting for these non-existing threads to release table locks
cause the dead lock.
In general, the design of Log_to_csv_event_handler does not fit into the
general architecture of the server, so that the concept of general_log_thd
and slow_log_thd has to be abandoned:
- this implementation does not work with table locking
- it will not work with commands like SHOW PROCESSLIST
- having the log tables always opened does not integrate well with DDL
operations / FLUSH TABLES / SET GLOBAL READ_ONLY
With this patch, the fundamental design of the LOGGER has been changed to:
- always open and close a log table when writing a log
- remove totally the usage of fake THD objects
- clarify how locking of log tables is implemented in general.
See WL#3984 for details related to the new locking design.
Additional changes (misc bugs exposed and fixed):
1)
mysqldump which would ignore some tables in dump_all_tables_in_db(),
but forget to ignore the same in dump_all_views_in_db().
2)
mysqldump would also issue an empty "LOCK TABLE" command when all the tables
to lock are to be ignored (numrows == 0), instead of not issuing the query.
3)
Internal errors handlers could intercept errors but not warnings
(see sql_error.cc).
4)
Implementing a nested call to open tables, for the performance schema tables,
exposed an existing bug in remove_table_from_cache(), which would perform:
in_use->some_tables_deleted=1;
against another thread, without any consideration about thread locking.
This call inside remove_table_from_cache() was not required anyway,
since calling mysql_lock_abort() takes care of aborting -- cleanly -- threads
that might hold a lock on a table.
This line (in_use->some_tables_deleted=1) has been removed.
2007-07-27 08:31:06 +02:00
|
|
|
thd->restore_backup_open_tables_state(backup);
|
|
|
|
}
|
|
|
|
|
2007-08-15 17:08:44 +02:00
|
|
|
/**
|
|
|
|
@} (end of group Data_Dictionary)
|
|
|
|
*/
|