2017-04-18 20:43:20 +02:00
|
|
|
/******************************************************
|
2017-10-25 08:06:45 +02:00
|
|
|
MariaBackup: hot backup tool for InnoDB
|
2017-04-19 15:09:03 +02:00
|
|
|
(c) 2009-2017 Percona LLC and/or its affiliates
|
2017-04-18 20:43:20 +02:00
|
|
|
Originally Created 3/3/2009 Yasufumi Kinoshita
|
|
|
|
Written by Alexey Kopytov, Aleksandr Kuzminsky, Stewart Smith, Vadim Tkachenko,
|
|
|
|
Yasufumi Kinoshita, Ignacio Nin and Baron Schwartz.
|
2022-01-27 15:17:40 +01:00
|
|
|
(c) 2017, 2022, MariaDB Corporation.
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
Portions written by Marko Mäkelä.
|
2017-04-18 20:43:20 +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
|
|
|
|
the Free Software Foundation; version 2 of the License.
|
|
|
|
|
|
|
|
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.
|
|
|
|
|
|
|
|
You should have received a copy of the GNU General Public License
|
|
|
|
along with this program; if not, write to the Free Software
|
2019-05-11 21:19:05 +02:00
|
|
|
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1335 USA
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
*******************************************************
|
|
|
|
|
|
|
|
This file incorporates work covered by the following copyright and
|
|
|
|
permission notice:
|
|
|
|
|
|
|
|
Copyright (c) 2000, 2011, MySQL AB & Innobase Oy. All Rights Reserved.
|
|
|
|
|
|
|
|
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 the Free Software
|
|
|
|
Foundation; version 2 of the License.
|
|
|
|
|
|
|
|
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.
|
|
|
|
|
|
|
|
You should have received a copy of the GNU General Public License along with
|
2019-05-11 18:25:02 +02:00
|
|
|
this program; if not, write to the Free Software Foundation, Inc., 51 Franklin
|
|
|
|
Street, Fifth Floor, Boston, MA 02110-1335 USA
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
*******************************************************/
|
|
|
|
|
|
|
|
//#define XTRABACKUP_TARGET_IS_PLUGIN
|
|
|
|
|
2017-06-19 05:34:38 +02:00
|
|
|
#include <my_global.h>
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#include <my_config.h>
|
|
|
|
#include <unireg.h>
|
2017-04-18 20:43:20 +02:00
|
|
|
#include <mysql_version.h>
|
|
|
|
#include <my_base.h>
|
|
|
|
#include <my_getopt.h>
|
|
|
|
#include <mysql_com.h>
|
|
|
|
#include <my_default.h>
|
2022-05-06 09:49:35 +02:00
|
|
|
#include <scope.h>
|
2021-07-20 13:00:31 +02:00
|
|
|
#include <sql_class.h>
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
#include <fcntl.h>
|
|
|
|
#include <string.h>
|
|
|
|
|
|
|
|
#ifdef __linux__
|
|
|
|
# include <sys/prctl.h>
|
2021-03-19 10:31:27 +01:00
|
|
|
# include <sys/resource.h>
|
2017-04-18 20:43:20 +02:00
|
|
|
#endif
|
|
|
|
|
2021-02-03 20:44:34 +01:00
|
|
|
#ifdef __APPLE__
|
|
|
|
# include "libproc.h"
|
|
|
|
#endif
|
|
|
|
|
2021-03-19 10:31:27 +01:00
|
|
|
#ifdef __FreeBSD__
|
|
|
|
# include <sys/sysctl.h>
|
|
|
|
#endif
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
#include <btr0sea.h>
|
|
|
|
#include <lock0lock.h>
|
|
|
|
#include <log0recv.h>
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#include <log0crypt.h>
|
2017-04-18 20:43:20 +02:00
|
|
|
#include <row0mysql.h>
|
|
|
|
#include <row0quiesce.h>
|
|
|
|
#include <srv0start.h>
|
2018-03-11 22:34:23 +01:00
|
|
|
#include "trx0sys.h"
|
2017-04-18 20:43:20 +02:00
|
|
|
#include <buf0dblwr.h>
|
2020-04-02 23:43:09 +02:00
|
|
|
#include "ha_innodb.h"
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
#include <list>
|
2017-04-18 20:43:20 +02:00
|
|
|
#include <sstream>
|
|
|
|
#include <set>
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
#include <fstream>
|
2017-04-18 20:43:20 +02:00
|
|
|
#include <mysql.h>
|
|
|
|
|
|
|
|
#define G_PTR uchar*
|
|
|
|
|
|
|
|
#include "common.h"
|
|
|
|
#include "datasink.h"
|
|
|
|
|
|
|
|
#include "xb_regex.h"
|
|
|
|
#include "fil_cur.h"
|
|
|
|
#include "write_filt.h"
|
|
|
|
#include "xtrabackup.h"
|
|
|
|
#include "ds_buffer.h"
|
|
|
|
#include "ds_tmpfile.h"
|
|
|
|
#include "xbstream.h"
|
|
|
|
#include "changed_page_bitmap.h"
|
|
|
|
#include "read_filt.h"
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#include "backup_wsrep.h"
|
2017-04-18 20:43:20 +02:00
|
|
|
#include "innobackupex.h"
|
|
|
|
#include "backup_mysql.h"
|
|
|
|
#include "backup_copy.h"
|
|
|
|
#include "backup_mysql.h"
|
2021-10-22 10:14:57 +02:00
|
|
|
#include "xb_plugin.h"
|
2017-04-18 21:05:57 +02:00
|
|
|
#include <sql_plugin.h>
|
|
|
|
#include <srv0srv.h>
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#include <log.h>
|
2019-12-11 11:22:03 +01:00
|
|
|
#include <derror.h>
|
2019-10-29 22:37:12 +01:00
|
|
|
#include <thr_timer.h>
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
#include "backup_debug.h"
|
|
|
|
|
|
|
|
#define MB_CORRUPTED_PAGES_FILE "innodb_corrupted_pages"
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
int sys_var_init();
|
|
|
|
|
|
|
|
/* === xtrabackup specific options === */
|
|
|
|
char xtrabackup_real_target_dir[FN_REFLEN] = "./xtrabackup_backupfiles/";
|
|
|
|
char *xtrabackup_target_dir= xtrabackup_real_target_dir;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
static my_bool xtrabackup_version;
|
2018-06-15 13:09:15 +02:00
|
|
|
static my_bool verbose;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
my_bool xtrabackup_backup;
|
|
|
|
my_bool xtrabackup_prepare;
|
|
|
|
my_bool xtrabackup_copy_back;
|
|
|
|
my_bool xtrabackup_move_back;
|
|
|
|
my_bool xtrabackup_decrypt_decompress;
|
|
|
|
my_bool xtrabackup_print_param;
|
2020-06-08 10:45:56 +02:00
|
|
|
my_bool xtrabackup_mysqld_args;
|
2020-06-23 17:07:03 +02:00
|
|
|
my_bool xtrabackup_help;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
|
|
|
my_bool xtrabackup_export;
|
|
|
|
|
|
|
|
longlong xtrabackup_use_memory;
|
|
|
|
|
2017-08-21 19:16:12 +02:00
|
|
|
uint opt_protocol;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
long xtrabackup_throttle; /* 0:unlimited */
|
|
|
|
static lint io_ticket;
|
2020-12-04 15:18:04 +01:00
|
|
|
static mysql_cond_t wait_throttle;
|
|
|
|
static mysql_cond_t log_copying_stop;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
|
|
|
char *xtrabackup_incremental;
|
2017-04-18 20:43:20 +02:00
|
|
|
lsn_t incremental_lsn;
|
|
|
|
lsn_t incremental_to_lsn;
|
|
|
|
lsn_t incremental_last_lsn;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
xb_page_bitmap *changed_page_bitmap;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
char *xtrabackup_incremental_basedir; /* for --backup */
|
|
|
|
char *xtrabackup_extra_lsndir; /* for --backup with --extra-lsndir */
|
|
|
|
char *xtrabackup_incremental_dir; /* for --prepare */
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
char xtrabackup_real_incremental_basedir[FN_REFLEN];
|
|
|
|
char xtrabackup_real_extra_lsndir[FN_REFLEN];
|
|
|
|
char xtrabackup_real_incremental_dir[FN_REFLEN];
|
|
|
|
|
2017-09-08 00:53:21 +02:00
|
|
|
|
2017-04-18 21:05:57 +02:00
|
|
|
char *xtrabackup_tmpdir;
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
char *xtrabackup_tables;
|
|
|
|
char *xtrabackup_tables_file;
|
|
|
|
char *xtrabackup_tables_exclude;
|
2018-06-07 16:13:54 +02:00
|
|
|
char *xb_rocksdb_datadir;
|
|
|
|
my_bool xb_backup_rocksdb = 1;
|
2017-04-19 15:09:03 +02:00
|
|
|
|
|
|
|
typedef std::list<regex_t> regex_list_t;
|
|
|
|
static regex_list_t regex_include_list;
|
|
|
|
static regex_list_t regex_exclude_list;
|
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
static hash_table_t tables_include_hash;
|
|
|
|
static hash_table_t tables_exclude_hash;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
char *xtrabackup_databases = NULL;
|
|
|
|
char *xtrabackup_databases_file = NULL;
|
2017-04-19 15:09:03 +02:00
|
|
|
char *xtrabackup_databases_exclude = NULL;
|
2020-06-18 11:26:28 +02:00
|
|
|
static hash_table_t databases_include_hash;
|
|
|
|
static hash_table_t databases_exclude_hash;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
static hash_table_t inc_dir_tables_hash;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-25691: Simplify handlerton::drop_database for InnoDB
The implementation of handlerton::drop_database in InnoDB is
unnecessarily complex. The minimal implementation should check
that no conflicting locks or references exist on the tables,
delete all table metadata in a single transaction, and finally
delete the tablespaces.
Note: DROP DATABASE will delete each individual table that the
SQL layer knows about, one table per transaction.
The handlerton::drop_database is basically a final cleanup step
for removing any garbage that could have been left behind
in InnoDB due to some bug, or not having atomic DDL in the past.
hash_node_t: Remove. Use the proper data type name in pointers.
dict_drop_index_tree(): Do not take the table as a parameter.
Instead, return the tablespace ID if the tablespace should be dropped
(we are dropping a clustered index tree).
fil_delete_tablespace(), fil_system_t::detach(): Return a single
detached file handle. Multi-file tablespaces cannot be deleted
via this interface.
ha_innobase::delete_table(): Remove a work-around for non-atomic DDL
and do not try to drop tables with similar-looking name.
innodb_drop_database(): Complete rewrite.
innobase_drop_database(), dict_get_first_table_name_in_db(),
row_drop_database_for_mysql(), drop_all_foreign_keys_in_db(): Remove.
row_purge_remove_clust_if_poss_low(), row_undo_ins_remove_clust_rec():
If the tablespace is to be deleted, try to evict the table definition
from the cache. Failing that, set dict_table_t::space to nullptr.
lock_release_on_rollback(): On the rollback of CREATE TABLE, release all
locks that the transaction had on the table, to avoid heap-use-after-free.
2021-05-18 11:53:40 +02:00
|
|
|
struct xb_filter_entry_t{
|
2017-04-18 20:43:20 +02:00
|
|
|
char* name;
|
|
|
|
ibool has_tables;
|
MDEV-25691: Simplify handlerton::drop_database for InnoDB
The implementation of handlerton::drop_database in InnoDB is
unnecessarily complex. The minimal implementation should check
that no conflicting locks or references exist on the tables,
delete all table metadata in a single transaction, and finally
delete the tablespaces.
Note: DROP DATABASE will delete each individual table that the
SQL layer knows about, one table per transaction.
The handlerton::drop_database is basically a final cleanup step
for removing any garbage that could have been left behind
in InnoDB due to some bug, or not having atomic DDL in the past.
hash_node_t: Remove. Use the proper data type name in pointers.
dict_drop_index_tree(): Do not take the table as a parameter.
Instead, return the tablespace ID if the tablespace should be dropped
(we are dropping a clustered index tree).
fil_delete_tablespace(), fil_system_t::detach(): Return a single
detached file handle. Multi-file tablespaces cannot be deleted
via this interface.
ha_innobase::delete_table(): Remove a work-around for non-atomic DDL
and do not try to drop tables with similar-looking name.
innodb_drop_database(): Complete rewrite.
innobase_drop_database(), dict_get_first_table_name_in_db(),
row_drop_database_for_mysql(), drop_all_foreign_keys_in_db(): Remove.
row_purge_remove_clust_if_poss_low(), row_undo_ins_remove_clust_rec():
If the tablespace is to be deleted, try to evict the table definition
from the cache. Failing that, set dict_table_t::space to nullptr.
lock_release_on_rollback(): On the rollback of CREATE TABLE, release all
locks that the transaction had on the table, to avoid heap-use-after-free.
2021-05-18 11:53:40 +02:00
|
|
|
xb_filter_entry_t *name_hash;
|
2017-04-18 20:43:20 +02:00
|
|
|
};
|
|
|
|
|
2023-01-10 13:42:50 +01:00
|
|
|
/** whether log_copying_thread() is active; protected by recv_sys.mutex */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
static bool log_copying_running;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
int xtrabackup_parallel;
|
|
|
|
|
|
|
|
char *xtrabackup_stream_str = NULL;
|
|
|
|
xb_stream_fmt_t xtrabackup_stream_fmt = XB_STREAM_FMT_NONE;
|
|
|
|
ibool xtrabackup_stream = FALSE;
|
|
|
|
|
|
|
|
const char *xtrabackup_compress_alg = NULL;
|
2019-01-14 22:28:23 +01:00
|
|
|
uint xtrabackup_compress = FALSE;
|
2017-04-18 20:43:20 +02:00
|
|
|
uint xtrabackup_compress_threads;
|
|
|
|
ulonglong xtrabackup_compress_chunk_size = 0;
|
|
|
|
|
|
|
|
/* sleep interval beetween log copy iterations in log copying thread
|
|
|
|
in milliseconds (default is 1 second) */
|
|
|
|
ulint xtrabackup_log_copy_interval = 1000;
|
2017-04-18 21:05:57 +02:00
|
|
|
static ulong max_buf_pool_modified_pct;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* Ignored option (--log) for MySQL option compatibility */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
static char* log_ignored_opt;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2018-05-08 16:08:08 +02:00
|
|
|
|
|
|
|
extern my_bool opt_use_ssl;
|
MDEV-14101 Provide an option to select TLS protocol version
Server and command line tools now support option --tls_version to specify the
TLS version between client and server. Valid values are TLSv1.0, TLSv1.1, TLSv1.2, TLSv1.3
or a combination of them. E.g.
--tls_version=TLSv1.3
--tls_version=TLSv1.2,TLSv1.3
In case there is a gap between versions, the lowest version will be used:
--tls_version=TLSv1.1,TLSv1.3 -> Only TLSv1.1 will be available.
If the used TLS library doesn't support the specified TLS version, it will use
the default configuration.
Limitations:
SSLv3 is not supported. The default configuration doesn't support TLSv1.0 anymore.
TLSv1.3 protocol currently is only supported by OpenSSL 1.1.0 (client and server) and
GnuTLS 3.6.5 (client only).
Overview of TLS implementations and protocols
Server:
+-----------+-----------------------------------------+
| Library | Supported TLS versions |
+-----------+-----------------------------------------+
| WolfSSL | TLSv1.1, TLSv1,2 |
+-----------+-----------------------------------------+
| OpenSSL | (TLSv1.0), TLSv1.1, TLSv1,2, TLSv1.3 |
+-----------+-----------------------------------------+
| LibreSSL | (TLSv1.0), TLSv1.1, TLSv1,2, TLSv1.3 |
+-----------+-----------------------------------------+
Client (MariaDB Connector/C)
+-----------+-----------------------------------------+
| Library | Supported TLS versions |
+-----------+-----------------------------------------+
| GnuTLS | (TLSv1.0), TLSv1.1, TLSv1.2, TLSv1.3 |
+-----------+-----------------------------------------+
| Schannel | (TLSv1.0), TLSv1.1, TLSv1.2 |
+-----------+-----------------------------------------+
| OpenSSL | (TLSv1.0), TLSv1.1, TLSv1,2, TLSv1.3 |
+-----------+-----------------------------------------+
| LibreSSL | (TLSv1.0), TLSv1.1, TLSv1,2, TLSv1.3 |
+-----------+-----------------------------------------+
2019-06-11 12:44:16 +02:00
|
|
|
extern char *opt_tls_version;
|
2018-05-08 16:08:08 +02:00
|
|
|
my_bool opt_ssl_verify_server_cert;
|
2018-12-20 14:31:18 +01:00
|
|
|
my_bool opt_extended_validation;
|
2019-01-23 13:00:12 +01:00
|
|
|
my_bool opt_encrypted_backup;
|
2018-05-08 16:08:08 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* === metadata of backup === */
|
|
|
|
#define XTRABACKUP_METADATA_FILENAME "xtrabackup_checkpoints"
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
char metadata_type[30] = ""; /*[full-backuped|log-applied|incremental]*/
|
2018-06-07 13:11:55 +02:00
|
|
|
static lsn_t metadata_from_lsn;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
lsn_t metadata_to_lsn;
|
2018-06-07 13:11:55 +02:00
|
|
|
static lsn_t metadata_last_lsn;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
static ds_file_t* dst_log_file;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
static char mysql_data_home_buff[2];
|
|
|
|
|
|
|
|
const char *defaults_group = "mysqld";
|
|
|
|
|
|
|
|
/* === static parameters in ha_innodb.cc */
|
|
|
|
|
|
|
|
#define HA_INNOBASE_ROWS_IN_TABLE 10000 /* to get optimization right */
|
|
|
|
#define HA_INNOBASE_RANGE_COUNT 100
|
|
|
|
|
|
|
|
/* The default values for the following, type long or longlong, start-up
|
|
|
|
parameters are declared in mysqld.cc: */
|
|
|
|
|
|
|
|
long innobase_buffer_pool_awe_mem_mb = 0;
|
|
|
|
long innobase_file_io_threads = 4;
|
2020-03-25 15:09:14 +01:00
|
|
|
ulong innobase_read_io_threads = 4;
|
|
|
|
ulong innobase_write_io_threads = 4;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2023-01-02 08:43:59 +01:00
|
|
|
/** Store the failed read of undo tablespace ids. Protected by
|
2023-01-10 13:42:50 +01:00
|
|
|
recv_sys.mutex. */
|
2023-01-02 08:43:59 +01:00
|
|
|
static std::set<uint32_t> fail_undo_ids;
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
longlong innobase_page_size = (1LL << 14); /* 16KB */
|
2022-06-28 09:16:31 +02:00
|
|
|
char* innobase_buffer_pool_filename = NULL;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* The default values for the following char* start-up parameters
|
|
|
|
are determined in innobase_init below: */
|
|
|
|
|
2017-07-06 17:31:28 +02:00
|
|
|
static char* innobase_ignored_opt;
|
|
|
|
char* innobase_data_home_dir;
|
|
|
|
char* innobase_data_file_path;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
|
|
|
|
my_bool xtrabackup_incremental_force_scan = FALSE;
|
|
|
|
|
2021-04-01 11:33:59 +02:00
|
|
|
/*
|
|
|
|
* Ignore corrupt pages (disabled by default; used
|
|
|
|
* by "innobackupex" as a command line argument).
|
|
|
|
*/
|
|
|
|
ulong xtrabackup_innodb_force_recovery = 0;
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* The flushed lsn which is read from data files */
|
2017-05-30 10:30:43 +02:00
|
|
|
lsn_t flushed_lsn= 0;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
ulong xb_open_files_limit= 0;
|
2017-04-18 21:05:57 +02:00
|
|
|
char *xb_plugin_dir;
|
|
|
|
char *xb_plugin_load;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
my_bool xb_close_files;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* Datasinks */
|
|
|
|
ds_ctxt_t *ds_data = NULL;
|
|
|
|
ds_ctxt_t *ds_meta = NULL;
|
|
|
|
ds_ctxt_t *ds_redo = NULL;
|
|
|
|
|
|
|
|
static bool innobackupex_mode = false;
|
|
|
|
|
|
|
|
/* String buffer used by --print-param to accumulate server options as they are
|
|
|
|
parsed from the defaults file */
|
|
|
|
static std::ostringstream print_param_str;
|
|
|
|
|
|
|
|
/* Set of specified parameters */
|
|
|
|
std::set<std::string> param_set;
|
|
|
|
|
|
|
|
static ulonglong global_max_value;
|
|
|
|
|
|
|
|
extern "C" sig_handler handle_fatal_signal(int sig);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
extern LOGGER logger;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
my_bool opt_galera_info = FALSE;
|
|
|
|
my_bool opt_slave_info = FALSE;
|
|
|
|
my_bool opt_no_lock = FALSE;
|
|
|
|
my_bool opt_safe_slave_backup = FALSE;
|
|
|
|
my_bool opt_rsync = FALSE;
|
|
|
|
my_bool opt_force_non_empty_dirs = FALSE;
|
|
|
|
my_bool opt_noversioncheck = FALSE;
|
|
|
|
my_bool opt_no_backup_locks = FALSE;
|
|
|
|
my_bool opt_decompress = FALSE;
|
2017-10-25 16:17:21 +02:00
|
|
|
my_bool opt_remove_original;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
my_bool opt_log_innodb_page_corruption;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-09-11 18:45:36 +02:00
|
|
|
my_bool opt_lock_ddl_per_table = FALSE;
|
2018-12-14 23:36:21 +01:00
|
|
|
static my_bool opt_check_privileges;
|
2017-09-11 18:45:36 +02:00
|
|
|
|
2018-04-30 17:12:14 +02:00
|
|
|
extern const char *innodb_checksum_algorithm_names[];
|
|
|
|
extern TYPELIB innodb_checksum_algorithm_typelib;
|
|
|
|
extern const char *innodb_flush_method_names[];
|
|
|
|
extern TYPELIB innodb_flush_method_typelib;
|
MDEV-30136: Deprecate innodb_flush_method
We introduce the following settable Boolean global variables:
innodb_log_file_write_through: Whether writes to ib_logfile0 are
write-through (disabling any caching, as in O_SYNC or O_DSYNC).
innodb_data_file_write_through: Whether writes to any InnoDB data files
(including the temporary tablespace) are write-through.
innodb_data_file_buffering: Whether the file system cache is enabled
for InnoDB data files.
All these parameters are OFF by default, that is, the file system cache
will be disabled, but any hardware caching is enabled, that is,
explicit calls to fsync(), fdatasync() or similar functions are needed.
On systems that support FUA it may make sense to enable write-through,
to avoid extra system calls.
If the deprecated read-only start-up parameter is set to one of the
following values, then the values of the 4 Boolean flags (the above 3
plus innodb_log_file_buffering) will be set as follows:
O_DSYNC:
innodb_log_file_write_through=ON, innodb_data_file_write_through=ON,
innodb_data_file_buffering=OFF, and
(if supported) innodb_log_file_buffering=OFF.
fsync, littlesync, nosync, or (Microsoft Windows specific) normal:
innodb_log_file_write_through=OFF, innodb_data_file_write_through=OFF,
and innodb_data_file_buffering=ON.
Note: fsync() or fdatasync() will only be disabled if the separate
parameter debug_no_sync (in the code, my_disable_sync) is set.
In mariadb-backup, the parameter innodb_flush_method will be ignored.
The Boolean parameters can be modified by SET GLOBAL while the
server is running. This will require reopening the ib_logfile0
or all currently open InnoDB data files.
We will open files straight in O_DSYNC or O_SYNC mode when applicable.
Data files we will try to open straight in O_DIRECT mode when the
page size is at least 4096 bytes. For atomically creating data files,
we will invoke os_file_set_nocache() to enable O_DIRECT afterwards,
because O_DIRECT is not supported on some file systems. We will also
continue to invoke os_file_set_nocache() on ib_logfile0 when
innodb_log_file_buffering=OFF can be fulfilled.
For reopening the ib_logfile0, we use the same logic that was developed
for online log resizing and reused for updates of
innodb_log_file_buffering.
Reopening all data files is implemented in the new function
fil_space_t::reopen_all().
Reviewed by: Vladislav Vaintroub
Tested by: Matthias Leich
2022-12-14 13:43:32 +01:00
|
|
|
/** Ignored option */
|
|
|
|
static ulong innodb_flush_method;
|
2018-04-30 17:12:14 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
static const char *binlog_info_values[] = {"off", "lockless", "on", "auto",
|
|
|
|
NullS};
|
|
|
|
static TYPELIB binlog_info_typelib = {array_elements(binlog_info_values)-1, "",
|
|
|
|
binlog_info_values, NULL};
|
|
|
|
ulong opt_binlog_info;
|
|
|
|
|
2017-07-06 17:31:28 +02:00
|
|
|
char *opt_incremental_history_name;
|
|
|
|
char *opt_incremental_history_uuid;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-07-06 17:31:28 +02:00
|
|
|
char *opt_user;
|
2021-02-05 13:57:46 +01:00
|
|
|
const char *opt_password;
|
2017-07-06 17:31:28 +02:00
|
|
|
char *opt_host;
|
|
|
|
char *opt_defaults_group;
|
|
|
|
char *opt_socket;
|
|
|
|
uint opt_port;
|
|
|
|
char *opt_log_bin;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
const char *query_type_names[] = { "ALL", "UPDATE", "SELECT", NullS};
|
|
|
|
|
|
|
|
TYPELIB query_type_typelib= {array_elements(query_type_names) - 1, "",
|
|
|
|
query_type_names, NULL};
|
|
|
|
|
|
|
|
ulong opt_lock_wait_query_type;
|
|
|
|
ulong opt_kill_long_query_type;
|
|
|
|
|
|
|
|
uint opt_kill_long_queries_timeout = 0;
|
|
|
|
uint opt_lock_wait_timeout = 0;
|
|
|
|
uint opt_lock_wait_threshold = 0;
|
|
|
|
uint opt_debug_sleep_before_unlock = 0;
|
|
|
|
uint opt_safe_slave_backup_timeout = 0;
|
|
|
|
|
|
|
|
const char *opt_history = NULL;
|
|
|
|
|
|
|
|
|
2017-09-08 00:53:21 +02:00
|
|
|
char mariabackup_exe[FN_REFLEN];
|
|
|
|
char orig_argv1[FN_REFLEN];
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
pthread_cond_t scanned_lsn_cond;
|
|
|
|
|
2022-01-29 13:40:25 +01:00
|
|
|
/** Store the deferred tablespace name during --backup */
|
|
|
|
static std::set<std::string> defer_space_names;
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
typedef std::map<space_id_t,std::string> space_id_to_name_t;
|
|
|
|
|
|
|
|
struct ddl_tracker_t {
|
|
|
|
/** Tablspaces with their ID and name, as they were copied to backup.*/
|
|
|
|
space_id_to_name_t tables_in_backup;
|
|
|
|
/** Drop operations found in redo log. */
|
|
|
|
std::set<space_id_t> drops;
|
|
|
|
/* For DDL operation found in redo log, */
|
|
|
|
space_id_to_name_t id_to_name;
|
2022-01-29 13:40:25 +01:00
|
|
|
/** Deferred tablespaces with their ID and name which was
|
|
|
|
found in redo log of DDL operations */
|
|
|
|
space_id_to_name_t deferred_tables;
|
|
|
|
|
|
|
|
/** Insert the deferred tablespace id with the name */
|
|
|
|
void insert_defer_id(space_id_t space_id, std::string name)
|
|
|
|
{
|
|
|
|
auto it= defer_space_names.find(name);
|
|
|
|
if (it != defer_space_names.end())
|
|
|
|
{
|
|
|
|
deferred_tables[space_id]= name;
|
|
|
|
defer_space_names.erase(it);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Rename the deferred tablespace with new name */
|
|
|
|
void rename_defer(space_id_t space_id, std::string old_name,
|
|
|
|
std::string new_name)
|
|
|
|
{
|
|
|
|
if (deferred_tables.find(space_id) != deferred_tables.end())
|
|
|
|
deferred_tables[space_id] = new_name;
|
|
|
|
auto defer_end= defer_space_names.end();
|
|
|
|
auto defer= defer_space_names.find(old_name);
|
|
|
|
if (defer == defer_end)
|
|
|
|
defer= defer_space_names.find(new_name);
|
|
|
|
|
|
|
|
if (defer != defer_end)
|
|
|
|
{
|
|
|
|
deferred_tables[space_id]= new_name;
|
|
|
|
defer_space_names.erase(defer);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Delete the deferred tablespace */
|
|
|
|
void delete_defer(space_id_t space_id, std::string name)
|
|
|
|
{
|
|
|
|
deferred_tables.erase(space_id);
|
|
|
|
defer_space_names.erase(name);
|
|
|
|
}
|
2018-08-09 16:06:52 +02:00
|
|
|
};
|
|
|
|
|
2018-08-16 15:10:18 +02:00
|
|
|
static ddl_tracker_t ddl_tracker;
|
2018-08-09 16:06:52 +02:00
|
|
|
|
2023-01-02 08:43:59 +01:00
|
|
|
/** Store the space ids of truncated undo log tablespaces. Protected
|
|
|
|
by recv_sys.mutex */
|
|
|
|
static std::set<uint32_t> undo_trunc_ids;
|
|
|
|
|
2022-01-29 13:40:25 +01:00
|
|
|
/** Stores the space ids of page0 INIT_PAGE redo records. It is
|
|
|
|
used to indicate whether the given deferred tablespace can
|
|
|
|
be reconstructed. */
|
|
|
|
static std::set<space_id_t> first_page_init_ids;
|
|
|
|
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
// Convert non-null terminated filename to space name
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
static std::string filename_to_spacename(const void *filename, size_t len);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
|
|
|
|
CorruptedPages::CorruptedPages() { ut_a(!pthread_mutex_init(&m_mutex, NULL)); }
|
|
|
|
|
|
|
|
CorruptedPages::~CorruptedPages() { ut_a(!pthread_mutex_destroy(&m_mutex)); }
|
|
|
|
|
2021-07-22 10:22:47 +02:00
|
|
|
void CorruptedPages::add_page_no_lock(const char *space_name,
|
|
|
|
page_id_t page_id,
|
2020-12-02 17:29:49 +01:00
|
|
|
bool convert_space_name)
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
{
|
2021-07-22 10:22:47 +02:00
|
|
|
space_info_t &space_info = m_spaces[page_id.space()];
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
if (space_info.space_name.empty())
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
space_info.space_name= convert_space_name
|
|
|
|
? filename_to_spacename(space_name, strlen(space_name))
|
|
|
|
: space_name;
|
2021-07-22 10:22:47 +02:00
|
|
|
(void)space_info.pages.insert(page_id.page_no());
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
}
|
|
|
|
|
2021-07-22 10:22:47 +02:00
|
|
|
void CorruptedPages::add_page(const char *file_name, page_id_t page_id)
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
{
|
2021-07-22 10:22:47 +02:00
|
|
|
pthread_mutex_lock(&m_mutex);
|
|
|
|
add_page_no_lock(file_name, page_id, true);
|
|
|
|
pthread_mutex_unlock(&m_mutex);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
}
|
|
|
|
|
2021-07-22 10:22:47 +02:00
|
|
|
bool CorruptedPages::contains(page_id_t page_id) const
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
{
|
|
|
|
bool result = false;
|
|
|
|
ut_a(!pthread_mutex_lock(&m_mutex));
|
2021-07-22 10:22:47 +02:00
|
|
|
container_t::const_iterator space_it= m_spaces.find(page_id.space());
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
if (space_it != m_spaces.end())
|
2021-07-22 10:22:47 +02:00
|
|
|
result = space_it->second.pages.count(page_id.page_no());
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
ut_a(!pthread_mutex_unlock(&m_mutex));
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
2021-07-22 10:22:47 +02:00
|
|
|
void CorruptedPages::drop_space(uint32_t space_id)
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
{
|
|
|
|
ut_a(!pthread_mutex_lock(&m_mutex));
|
|
|
|
m_spaces.erase(space_id);
|
|
|
|
ut_a(!pthread_mutex_unlock(&m_mutex));
|
|
|
|
}
|
|
|
|
|
2021-07-22 10:22:47 +02:00
|
|
|
void CorruptedPages::rename_space(uint32_t space_id,
|
|
|
|
const std::string &new_name)
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
{
|
|
|
|
ut_a(!pthread_mutex_lock(&m_mutex));
|
|
|
|
container_t::iterator space_it = m_spaces.find(space_id);
|
|
|
|
if (space_it != m_spaces.end())
|
|
|
|
space_it->second.space_name = new_name;
|
|
|
|
ut_a(!pthread_mutex_unlock(&m_mutex));
|
|
|
|
}
|
|
|
|
|
|
|
|
bool CorruptedPages::print_to_file(const char *filename) const
|
|
|
|
{
|
|
|
|
std::ostringstream out;
|
|
|
|
ut_a(!pthread_mutex_lock(&m_mutex));
|
|
|
|
if (!m_spaces.size())
|
|
|
|
{
|
|
|
|
ut_a(!pthread_mutex_unlock(&m_mutex));
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
for (container_t::const_iterator space_it=
|
|
|
|
m_spaces.begin();
|
|
|
|
space_it != m_spaces.end(); ++space_it)
|
|
|
|
{
|
|
|
|
out << space_it->second.space_name << " " << space_it->first << "\n";
|
|
|
|
bool first_page_no= true;
|
2020-12-02 17:29:49 +01:00
|
|
|
for (std::set<unsigned>::const_iterator page_it=
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
space_it->second.pages.begin();
|
|
|
|
page_it != space_it->second.pages.end(); ++page_it)
|
|
|
|
if (first_page_no)
|
|
|
|
{
|
|
|
|
out << *page_it;
|
|
|
|
first_page_no= false;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
out << " " << *page_it;
|
|
|
|
out << "\n";
|
|
|
|
}
|
|
|
|
ut_a(!pthread_mutex_unlock(&m_mutex));
|
|
|
|
if (xtrabackup_backup)
|
|
|
|
return backup_file_print_buf(filename, out.str().c_str(),
|
2020-12-01 16:15:53 +01:00
|
|
|
static_cast<int>(out.str().size()));
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
std::ofstream outfile;
|
|
|
|
outfile.open(filename);
|
|
|
|
if (!outfile.is_open())
|
|
|
|
die("Can't open %s, error number: %d, error message: %s", filename, errno,
|
|
|
|
strerror(errno));
|
|
|
|
outfile << out.str();
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
void CorruptedPages::read_from_file(const char *file_name)
|
|
|
|
{
|
|
|
|
MY_STAT mystat;
|
|
|
|
if (!my_stat(file_name, &mystat, MYF(0)))
|
|
|
|
return;
|
|
|
|
std::ifstream infile;
|
|
|
|
infile.open(file_name);
|
|
|
|
if (!infile.is_open())
|
|
|
|
die("Can't open %s, error number: %d, error message: %s", file_name, errno,
|
|
|
|
strerror(errno));
|
|
|
|
std::string line;
|
|
|
|
std::string space_name;
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t space_id;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
ulint line_number= 0;
|
|
|
|
while (std::getline(infile, line))
|
|
|
|
{
|
|
|
|
++line_number;
|
|
|
|
std::istringstream iss(line);
|
|
|
|
if (line_number & 1) {
|
|
|
|
if (!(iss >> space_name))
|
|
|
|
die("Can't parse space name from corrupted pages file at "
|
|
|
|
"line " ULINTPF,
|
|
|
|
line_number);
|
|
|
|
if (!(iss >> space_id))
|
|
|
|
die("Can't parse space id from corrupted pages file at line " ULINTPF,
|
|
|
|
line_number);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2020-12-02 17:29:49 +01:00
|
|
|
std::istringstream iss(line);
|
|
|
|
unsigned page_no;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
while ((iss >> page_no))
|
2021-07-22 10:22:47 +02:00
|
|
|
add_page_no_lock(space_name.c_str(), {space_id, page_no}, false);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
if (!iss.eof())
|
|
|
|
die("Corrupted pages file parse error on line number " ULINTPF,
|
|
|
|
line_number);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
bool CorruptedPages::empty() const
|
|
|
|
{
|
|
|
|
ut_a(!pthread_mutex_lock(&m_mutex));
|
|
|
|
bool result= !m_spaces.size();
|
|
|
|
ut_a(!pthread_mutex_unlock(&m_mutex));
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void xb_load_single_table_tablespace(const std::string &space_name,
|
2021-05-17 17:12:33 +02:00
|
|
|
bool set_size,
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t defer_space_id=0);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
static void xb_data_files_close();
|
2020-12-01 13:55:46 +01:00
|
|
|
static fil_space_t* fil_space_get_by_name(const char* name);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
|
|
|
|
void CorruptedPages::zero_out_free_pages()
|
|
|
|
{
|
|
|
|
container_t non_free_pages;
|
2020-12-02 17:29:49 +01:00
|
|
|
byte *zero_page=
|
|
|
|
static_cast<byte *>(aligned_malloc(srv_page_size, srv_page_size));
|
2020-12-01 13:55:46 +01:00
|
|
|
memset(zero_page, 0, srv_page_size);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
|
|
|
|
ut_a(!pthread_mutex_lock(&m_mutex));
|
|
|
|
for (container_t::const_iterator space_it= m_spaces.begin();
|
|
|
|
space_it != m_spaces.end(); ++space_it)
|
|
|
|
{
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t space_id = space_it->first;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
const std::string &space_name = space_it->second.space_name;
|
|
|
|
// There is no need to close tablespaces explixitly as they will be closed
|
|
|
|
// in innodb_shutdown().
|
|
|
|
xb_load_single_table_tablespace(space_name, false);
|
2020-12-02 17:29:49 +01:00
|
|
|
fil_space_t *space = fil_space_t::get(space_id);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
if (!space)
|
|
|
|
die("Can't find space object for space name %s to check corrupted page",
|
|
|
|
space_name.c_str());
|
2020-12-02 17:29:49 +01:00
|
|
|
for (std::set<unsigned>::const_iterator page_it=
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
space_it->second.pages.begin();
|
|
|
|
page_it != space_it->second.pages.end(); ++page_it)
|
|
|
|
{
|
MDEV-13542: Crashing on corrupted page is unhelpful
The approach to handling corruption that was chosen by Oracle in
commit 177d8b0c125b841c0650d27d735e3b87509dc286
is not really useful. Not only did it actually fail to prevent InnoDB
from crashing, but it is making things worse by blocking attempts to
rescue data from or rebuild a partially readable table.
We will try to prevent crashes in a different way: by propagating
errors up the call stack. We will never mark the clustered index
persistently corrupted, so that data recovery may be attempted by
reading from the table, or by rebuilding the table.
This should also fix MDEV-13680 (crash on btr_page_alloc() failure);
it was extensively tested with innodb_file_per_table=0 and a
non-autoextend system tablespace.
We should now avoid crashes in many cases, such as when a page
cannot be read or allocated, or an inconsistency is detected when
attempting to update multiple pages. We will not crash on double-free,
such as on the recovery of DDL in system tablespace in case something
was corrupted.
Crashes on corrupted data are still possible. The fault injection mechanism
that is introduced in the subsequent commit may help catch more of them.
buf_page_import_corrupt_failure: Remove the fault injection, and instead
corrupt some pages using Perl code in the tests.
btr_cur_pessimistic_insert(): Always reserve extents (except for the
change buffer), in order to prevent a subsequent allocation failure.
btr_pcur_open_at_rnd_pos(): Merged to the only caller ibuf_merge_pages().
btr_assert_not_corrupted(), btr_corruption_report(): Remove.
Similar checks are already part of btr_block_get().
FSEG_MAGIC_N_BYTES: Replaces FSEG_MAGIC_N_VALUE.
dict_hdr_get(), trx_rsegf_get_new(), trx_undo_page_get(),
trx_undo_page_get_s_latched(): Replaced with error-checking calls.
trx_rseg_t::get(mtr_t*): Replaces trx_rsegf_get().
trx_rseg_header_create(): Let the caller update the TRX_SYS page if needed.
trx_sys_create_sys_pages(): Merged with trx_sysf_create().
dict_check_tablespaces_and_store_max_id(): Do not access
DICT_HDR_MAX_SPACE_ID, because it was already recovered in dict_boot().
Merge dict_check_sys_tables() with this function.
dir_pathname(): Replaces os_file_make_new_pathname().
row_undo_ins_remove_sec(): Do not modify the undo page by adding
a terminating NUL byte to the record.
btr_decryption_failed(): Report decryption failures
dict_set_corrupted_by_space(), dict_set_encrypted_by_space(),
dict_set_corrupted_index_cache_only(): Remove.
dict_set_corrupted(): Remove the constant parameter dict_locked=false.
Never flag the clustered index corrupted in SYS_INDEXES, because
that would deny further access to the table. It might be possible to
repair the table by executing ALTER TABLE or OPTIMIZE TABLE, in case
no B-tree leaf page is corrupted.
dict_table_skip_corrupt_index(), dict_table_next_uncorrupted_index(),
row_purge_skip_uncommitted_virtual_index(): Remove, and refactor
the callers to read dict_index_t::type only once.
dict_table_is_corrupted(): Remove.
dict_index_t::is_btree(): Determine if the index is a valid B-tree.
BUF_GET_NO_LATCH, BUF_EVICT_IF_IN_POOL: Remove.
UNIV_BTR_DEBUG: Remove. Any inconsistency will no longer trigger
assertion failures, but error codes being returned.
buf_corrupt_page_release(): Replaced with a direct call to
buf_pool.corrupted_evict().
fil_invalid_page_access_msg(): Never crash on an invalid read;
let the caller of buf_page_get_gen() decide.
btr_pcur_t::restore_position(): Propagate failure status to the caller
by returning CORRUPTED.
opt_search_plan_for_table(): Simplify the code.
row_purge_del_mark(), row_purge_upd_exist_or_extern_func(),
row_undo_ins_remove_sec_rec(), row_undo_mod_upd_del_sec(),
row_undo_mod_del_mark_sec(): Avoid mem_heap_create()/mem_heap_free()
when no secondary indexes exist.
row_undo_mod_upd_exist_sec(): Simplify the code.
row_upd_clust_step(), dict_load_table_one(): Return DB_TABLE_CORRUPT
if the clustered index (and therefore the table) is corrupted, similar
to what we do in row_insert_for_mysql().
fut_get_ptr(): Replace with buf_page_get_gen() calls.
buf_page_get_gen(): Return nullptr and *err=DB_CORRUPTION
if the page is marked as freed. For other modes than
BUF_GET_POSSIBLY_FREED or BUF_PEEK_IF_IN_POOL this will
trigger a debug assertion failure. For BUF_GET_POSSIBLY_FREED,
we will return nullptr for freed pages, so that the callers
can be simplified. The purge of transaction history will be
a new user of BUF_GET_POSSIBLY_FREED, to avoid crashes on
corrupted data.
buf_page_get_low(): Never crash on a corrupted page, but simply
return nullptr.
fseg_page_is_allocated(): Replaces fseg_page_is_free().
fts_drop_common_tables(): Return an error if the transaction
was rolled back.
fil_space_t::set_corrupted(): Report a tablespace as corrupted if
it was not reported already.
fil_space_t::io(): Invoke fil_space_t::set_corrupted() to report
out-of-bounds page access or other errors.
Clean up mtr_t::page_lock()
buf_page_get_low(): Validate the page identifier (to check for
recently read corrupted pages) after acquiring the page latch.
buf_page_t::read_complete(): Flag uninitialized (all-zero) pages
with DB_FAIL. Return DB_PAGE_CORRUPTED on page number mismatch.
mtr_t::defer_drop_ahi(): Renamed from mtr_defer_drop_ahi().
recv_sys_t::free_corrupted_page(): Only set_corrupt_fs()
if any log records exist for the page. We do not mind if read-ahead
produces corrupted (or all-zero) pages that were not actually needed
during recovery.
recv_recover_page(): Return whether the operation succeeded.
recv_sys_t::recover_low(): Simplify the logic. Check for recovery error.
Thanks to Matthias Leich for testing this extensively and to the
authors of https://rr-project.org for making it easy to diagnose
and fix any failures that were found during the testing.
2022-06-06 13:03:22 +02:00
|
|
|
if (fseg_page_is_allocated(space, *page_it))
|
|
|
|
{
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
space_info_t &space_info = non_free_pages[space_id];
|
|
|
|
space_info.pages.insert(*page_it);
|
|
|
|
if (space_info.space_name.empty())
|
|
|
|
space_info.space_name = space_name;
|
2020-12-02 17:29:49 +01:00
|
|
|
msg("Error: corrupted page " UINT32PF
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
" of tablespace %s can not be fixed",
|
|
|
|
*page_it, space_name.c_str());
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2020-12-02 17:29:49 +01:00
|
|
|
space->reacquire();
|
|
|
|
auto err= space
|
|
|
|
->io(IORequest(IORequest::PUNCH_RANGE),
|
|
|
|
*page_it * srv_page_size, srv_page_size, zero_page)
|
|
|
|
.err;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
if (err != DB_SUCCESS)
|
2020-12-02 17:29:49 +01:00
|
|
|
die("Can't zero out corrupted page " UINT32PF " of tablespace %s",
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
*page_it, space_name.c_str());
|
2020-12-02 17:29:49 +01:00
|
|
|
msg("Corrupted page " UINT32PF
|
2022-03-03 03:18:43 +01:00
|
|
|
" of tablespace %s was successfully fixed.",
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
*page_it, space_name.c_str());
|
|
|
|
}
|
|
|
|
}
|
2020-12-11 08:05:26 +01:00
|
|
|
space->flush<true>();
|
2020-12-02 17:29:49 +01:00
|
|
|
space->release();
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
}
|
|
|
|
m_spaces.swap(non_free_pages);
|
|
|
|
ut_a(!pthread_mutex_unlock(&m_mutex));
|
2020-12-02 17:29:49 +01:00
|
|
|
aligned_free(zero_page);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* Simple datasink creation tracking...add datasinks in the reverse order you
|
|
|
|
want them destroyed. */
|
|
|
|
#define XTRABACKUP_MAX_DATASINKS 10
|
|
|
|
static ds_ctxt_t *datasinks[XTRABACKUP_MAX_DATASINKS];
|
|
|
|
static uint actual_datasinks = 0;
|
|
|
|
static inline
|
|
|
|
void
|
|
|
|
xtrabackup_add_datasink(ds_ctxt_t *ds)
|
|
|
|
{
|
|
|
|
xb_ad(actual_datasinks < XTRABACKUP_MAX_DATASINKS);
|
|
|
|
datasinks[actual_datasinks] = ds; actual_datasinks++;
|
|
|
|
}
|
|
|
|
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
typedef void (*process_single_tablespace_func_t)(const char *dirname,
|
|
|
|
const char *filname,
|
|
|
|
bool is_remote,
|
2021-05-17 17:12:33 +02:00
|
|
|
bool skip_node_page0,
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t defer_space_id);
|
2017-09-08 00:53:21 +02:00
|
|
|
static dberr_t enumerate_ibd_files(process_single_tablespace_func_t callback);
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* ======== Datafiles iterator ======== */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
struct datafiles_iter_t {
|
2021-03-19 23:31:01 +01:00
|
|
|
space_list_t::iterator space = fil_system.space_list.end();
|
|
|
|
fil_node_t *node = nullptr;
|
|
|
|
bool started = false;
|
|
|
|
std::mutex mutex;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
};
|
|
|
|
|
|
|
|
/* ======== Datafiles iterator ======== */
|
|
|
|
static
|
2017-04-18 20:43:20 +02:00
|
|
|
fil_node_t *
|
|
|
|
datafiles_iter_next(datafiles_iter_t *it)
|
|
|
|
{
|
|
|
|
fil_node_t *new_node;
|
|
|
|
|
2021-03-19 23:31:01 +01:00
|
|
|
std::lock_guard<std::mutex> _(it->mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (it->node == NULL) {
|
|
|
|
if (it->started)
|
|
|
|
goto end;
|
|
|
|
it->started = TRUE;
|
|
|
|
} else {
|
|
|
|
it->node = UT_LIST_GET_NEXT(chain, it->node);
|
|
|
|
if (it->node != NULL)
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
|
2021-03-19 23:31:01 +01:00
|
|
|
it->space = (it->space == fil_system.space_list.end()) ?
|
|
|
|
fil_system.space_list.begin() :
|
|
|
|
std::next(it->space);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2021-03-19 23:31:01 +01:00
|
|
|
while (it->space != fil_system.space_list.end() &&
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
(it->space->purpose != FIL_TYPE_TABLESPACE ||
|
2017-04-18 20:43:20 +02:00
|
|
|
UT_LIST_GET_LEN(it->space->chain) == 0))
|
2021-03-19 23:31:01 +01:00
|
|
|
++it->space;
|
|
|
|
if (it->space == fil_system.space_list.end())
|
2017-04-18 20:43:20 +02:00
|
|
|
goto end;
|
|
|
|
|
|
|
|
it->node = UT_LIST_GET_FIRST(it->space->chain);
|
|
|
|
|
|
|
|
end:
|
|
|
|
new_node = it->node;
|
|
|
|
|
|
|
|
return new_node;
|
|
|
|
}
|
|
|
|
|
2018-03-30 00:13:01 +02:00
|
|
|
#ifndef DBUG_OFF
|
|
|
|
struct dbug_thread_param_t
|
|
|
|
{
|
|
|
|
MYSQL *con;
|
|
|
|
const char *query;
|
|
|
|
int expect_err;
|
|
|
|
int expect_errno;
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
|
|
/* Thread procedure used in dbug_start_query_thread. */
|
2020-12-04 15:18:04 +01:00
|
|
|
static void *dbug_execute_in_new_connection(void *arg)
|
2018-03-30 00:13:01 +02:00
|
|
|
{
|
|
|
|
mysql_thread_init();
|
2020-12-04 15:18:04 +01:00
|
|
|
dbug_thread_param_t *par= static_cast<dbug_thread_param_t*>(arg);
|
2018-03-30 00:13:01 +02:00
|
|
|
int err = mysql_query(par->con, par->query);
|
|
|
|
int err_no = mysql_errno(par->con);
|
2018-11-19 20:43:04 +01:00
|
|
|
if(par->expect_err != err)
|
|
|
|
{
|
|
|
|
msg("FATAL: dbug_execute_in_new_connection : mysql_query '%s' returns %d, instead of expected %d",
|
|
|
|
par->query, err, par->expect_err);
|
|
|
|
_exit(1);
|
|
|
|
}
|
|
|
|
if (err && par->expect_errno && par->expect_errno != err_no)
|
|
|
|
{
|
|
|
|
msg("FATAL: dbug_execute_in_new_connection: mysql_query '%s' returns mysql_errno %d, instead of expected %d",
|
|
|
|
par->query, err_no, par->expect_errno);
|
|
|
|
_exit(1);
|
|
|
|
}
|
2018-03-30 00:13:01 +02:00
|
|
|
mysql_close(par->con);
|
|
|
|
mysql_thread_end();
|
|
|
|
delete par;
|
2020-12-04 15:18:04 +01:00
|
|
|
return nullptr;
|
2018-03-30 00:13:01 +02:00
|
|
|
}
|
|
|
|
|
2020-12-04 15:18:04 +01:00
|
|
|
static pthread_t dbug_alter_thread;
|
|
|
|
|
2018-03-30 00:13:01 +02:00
|
|
|
/*
|
|
|
|
Execute query from a new connection, in own thread.
|
|
|
|
|
|
|
|
@param query - query to be executed
|
|
|
|
@param wait_state - if not NULL, wait until query from new connection
|
|
|
|
reaches this state (value of column State in I_S.PROCESSLIST)
|
|
|
|
@param expected_err - if 0, query is supposed to finish successfully,
|
|
|
|
otherwise query should return error.
|
|
|
|
@param expected_errno - if not 0, and query finished with error,
|
|
|
|
expected mysql_errno()
|
|
|
|
*/
|
2020-12-04 15:18:04 +01:00
|
|
|
static void dbug_start_query_thread(
|
2018-03-30 00:13:01 +02:00
|
|
|
const char *query,
|
|
|
|
const char *wait_state,
|
|
|
|
int expected_err,
|
|
|
|
int expected_errno)
|
|
|
|
|
|
|
|
{
|
|
|
|
dbug_thread_param_t *par = new dbug_thread_param_t;
|
|
|
|
par->query = query;
|
|
|
|
par->expect_err = expected_err;
|
|
|
|
par->expect_errno = expected_errno;
|
|
|
|
par->con = xb_mysql_connect();
|
2020-12-04 15:18:04 +01:00
|
|
|
|
|
|
|
mysql_thread_create(0, &dbug_alter_thread, nullptr,
|
|
|
|
dbug_execute_in_new_connection, par);
|
2018-03-30 00:13:01 +02:00
|
|
|
|
|
|
|
if (!wait_state)
|
2020-12-04 15:18:04 +01:00
|
|
|
return;
|
2018-03-30 00:13:01 +02:00
|
|
|
|
|
|
|
char q[256];
|
|
|
|
snprintf(q, sizeof(q),
|
|
|
|
"SELECT 1 FROM INFORMATION_SCHEMA.PROCESSLIST where ID=%lu"
|
|
|
|
" AND Command='Query' AND State='%s'",
|
|
|
|
mysql_thread_id(par->con), wait_state);
|
|
|
|
for (;;) {
|
|
|
|
MYSQL_RES *result = xb_mysql_query(mysql_connection,q, true, true);
|
2018-09-21 10:18:59 +02:00
|
|
|
bool exists = mysql_fetch_row(result) != NULL;
|
|
|
|
mysql_free_result(result);
|
|
|
|
if (exists) {
|
2018-03-30 00:13:01 +02:00
|
|
|
goto end;
|
|
|
|
}
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Waiting for query '%s' on connection %lu to "
|
2018-03-30 00:13:01 +02:00
|
|
|
" reach state '%s'", query, mysql_thread_id(par->con),
|
|
|
|
wait_state);
|
|
|
|
my_sleep(1000);
|
|
|
|
}
|
|
|
|
end:
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("query '%s' on connection %lu reached state '%s'", query,
|
2018-03-30 00:13:01 +02:00
|
|
|
mysql_thread_id(par->con), wait_state);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2018-01-23 10:12:25 +01:00
|
|
|
void mdl_lock_all()
|
|
|
|
{
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
mdl_lock_init();
|
|
|
|
datafiles_iter_t it;
|
2018-01-23 10:12:25 +01:00
|
|
|
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
while (fil_node_t *node= datafiles_iter_next(&it))
|
|
|
|
{
|
|
|
|
const auto id= node->space->id;
|
|
|
|
if (const char *name= (fil_is_user_tablespace_id(id) &&
|
|
|
|
node->space->chain.start)
|
|
|
|
? node->space->chain.start->name : nullptr)
|
|
|
|
if (check_if_skip_table(filename_to_spacename(name,
|
|
|
|
strlen(name)).c_str()))
|
|
|
|
continue;
|
|
|
|
mdl_lock_table(id);
|
|
|
|
}
|
2018-01-23 10:12:25 +01:00
|
|
|
}
|
2018-03-13 10:49:30 +01:00
|
|
|
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
// Convert non-null terminated filename to space name
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
static std::string filename_to_spacename(const void *filename, size_t len)
|
2018-08-09 16:06:52 +02:00
|
|
|
{
|
|
|
|
// null- terminate filename
|
|
|
|
char *f = (char *)malloc(len + 1);
|
|
|
|
ut_a(f);
|
|
|
|
memcpy(f, filename, len);
|
|
|
|
f[len] = 0;
|
|
|
|
for (size_t i = 0; i < len; i++)
|
|
|
|
if (f[i] == '\\')
|
|
|
|
f[i] = '/';
|
|
|
|
char *p = strrchr(f, '.');
|
|
|
|
ut_a(p);
|
|
|
|
*p = 0;
|
|
|
|
char *table = strrchr(f, '/');
|
|
|
|
ut_a(table);
|
|
|
|
*table = 0;
|
|
|
|
char *db = strrchr(f, '/');
|
|
|
|
*table = '/';
|
2022-02-25 14:50:09 +01:00
|
|
|
std::string s(db ? db+1 : f);
|
2018-09-21 10:18:59 +02:00
|
|
|
free(f);
|
|
|
|
return s;
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
2018-03-13 10:49:30 +01:00
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
/** Report an operation to create, delete, or rename a file during backup.
|
|
|
|
@param[in] space_id tablespace identifier
|
2022-01-29 13:40:25 +01:00
|
|
|
@param[in] type redo log file operation type
|
2018-08-09 16:06:52 +02:00
|
|
|
@param[in] name file name (not NUL-terminated)
|
|
|
|
@param[in] len length of name, in bytes
|
|
|
|
@param[in] new_name new file name (NULL if not rename)
|
|
|
|
@param[in] new_len length of new_name, in bytes (0 if NULL) */
|
2022-02-04 14:11:46 +01:00
|
|
|
static void backup_file_op(uint32_t space_id, int type,
|
2018-08-09 16:06:52 +02:00
|
|
|
const byte* name, ulint len,
|
|
|
|
const byte* new_name, ulint new_len)
|
|
|
|
{
|
2018-03-13 10:49:30 +01:00
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
ut_ad(name);
|
|
|
|
ut_ad(len);
|
|
|
|
ut_ad(!new_name == !new_len);
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
mysql_mutex_assert_owner(&recv_sys.mutex);
|
2018-08-09 16:06:52 +02:00
|
|
|
|
2022-01-29 13:40:25 +01:00
|
|
|
switch(type) {
|
|
|
|
case FILE_CREATE:
|
|
|
|
{
|
|
|
|
std::string space_name = filename_to_spacename(name, len);
|
|
|
|
ddl_tracker.id_to_name[space_id] = space_name;
|
|
|
|
ddl_tracker.delete_defer(space_id, space_name);
|
2021-07-22 10:22:47 +02:00
|
|
|
msg("DDL tracking : create %u \"%.*s\"", space_id, int(len), name);
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
2022-01-29 13:40:25 +01:00
|
|
|
break;
|
|
|
|
case FILE_MODIFY:
|
|
|
|
ddl_tracker.insert_defer_id(
|
|
|
|
space_id, filename_to_spacename(name, len));
|
|
|
|
break;
|
|
|
|
case FILE_RENAME:
|
|
|
|
{
|
|
|
|
std::string new_space_name = filename_to_spacename(
|
|
|
|
new_name, new_len);
|
|
|
|
std::string old_space_name = filename_to_spacename(
|
|
|
|
name, len);
|
|
|
|
ddl_tracker.id_to_name[space_id] = new_space_name;
|
|
|
|
ddl_tracker.rename_defer(space_id, old_space_name,
|
|
|
|
new_space_name);
|
2021-07-22 10:22:47 +02:00
|
|
|
msg("DDL tracking : rename %u \"%.*s\",\"%.*s\"",
|
2018-08-09 16:06:52 +02:00
|
|
|
space_id, int(len), name, int(new_len), new_name);
|
2022-01-29 13:40:25 +01:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
case FILE_DELETE:
|
2018-08-09 16:06:52 +02:00
|
|
|
ddl_tracker.drops.insert(space_id);
|
2022-01-29 13:40:25 +01:00
|
|
|
ddl_tracker.delete_defer(
|
|
|
|
space_id, filename_to_spacename(name, len));
|
2021-07-22 10:22:47 +02:00
|
|
|
msg("DDL tracking : delete %u \"%.*s\"", space_id, int(len), name);
|
2022-01-29 13:40:25 +01:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
ut_ad(0);
|
|
|
|
break;
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
}
|
2018-03-13 10:49:30 +01:00
|
|
|
|
|
|
|
|
2018-09-14 10:35:18 +02:00
|
|
|
/*
|
|
|
|
This callback is called if DDL operation is detected,
|
|
|
|
at the end of backup
|
|
|
|
|
|
|
|
Normally, DDL operations are blocked due to FTWRL,
|
|
|
|
but in rare cases of --no-lock, they are not.
|
|
|
|
|
|
|
|
We will abort backup in this case.
|
|
|
|
*/
|
2022-02-04 14:11:46 +01:00
|
|
|
static void backup_file_op_fail(uint32_t space_id, int type,
|
2018-09-14 10:35:18 +02:00
|
|
|
const byte* name, ulint len,
|
|
|
|
const byte* new_name, ulint new_len)
|
|
|
|
{
|
2022-01-29 13:40:25 +01:00
|
|
|
bool fail = false;
|
|
|
|
switch(type) {
|
|
|
|
case FILE_CREATE:
|
2022-02-04 14:11:46 +01:00
|
|
|
msg("DDL tracking : create %u \"%.*s\"", space_id, int(len), name);
|
2022-01-29 13:40:25 +01:00
|
|
|
fail = !check_if_skip_table(
|
|
|
|
filename_to_spacename(name, len).c_str());
|
|
|
|
break;
|
|
|
|
case FILE_MODIFY:
|
|
|
|
break;
|
|
|
|
case FILE_RENAME:
|
2021-07-22 10:22:47 +02:00
|
|
|
msg("DDL tracking : rename %u \"%.*s\",\"%.*s\"",
|
2018-09-14 10:35:18 +02:00
|
|
|
space_id, int(len), name, int(new_len), new_name);
|
2022-01-29 13:40:25 +01:00
|
|
|
fail = !check_if_skip_table(
|
|
|
|
filename_to_spacename(name, len).c_str())
|
|
|
|
|| !check_if_skip_table(
|
|
|
|
filename_to_spacename(new_name, new_len).c_str());
|
|
|
|
break;
|
|
|
|
case FILE_DELETE:
|
|
|
|
fail = !check_if_skip_table(
|
|
|
|
filename_to_spacename(name, len).c_str());
|
2021-07-22 10:22:47 +02:00
|
|
|
msg("DDL tracking : delete %u \"%.*s\"", space_id, int(len), name);
|
2022-01-29 13:40:25 +01:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
ut_ad(0);
|
|
|
|
break;
|
2018-09-14 10:35:18 +02:00
|
|
|
}
|
2022-01-29 13:40:25 +01:00
|
|
|
|
2018-09-14 10:35:18 +02:00
|
|
|
if (fail) {
|
2019-02-14 20:18:34 +01:00
|
|
|
ut_a(opt_no_lock);
|
2019-01-15 22:47:54 +01:00
|
|
|
die("DDL operation detected in the late phase of backup."
|
2019-01-14 22:28:23 +01:00
|
|
|
"Backup is inconsistent. Remove --no-lock option to fix.");
|
2018-09-14 10:35:18 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-01-02 08:43:59 +01:00
|
|
|
static void backup_undo_trunc(uint32_t space_id)
|
|
|
|
{
|
|
|
|
undo_trunc_ids.insert(space_id);
|
|
|
|
}
|
2018-09-14 10:35:18 +02:00
|
|
|
|
2022-01-29 13:40:25 +01:00
|
|
|
/* Function to store the space id of page0 INIT_PAGE
|
|
|
|
@param space_id space id which has page0 init page */
|
2022-02-04 14:11:46 +01:00
|
|
|
static void backup_first_page_op(space_id_t space_id)
|
2022-01-29 13:40:25 +01:00
|
|
|
{
|
|
|
|
first_page_init_ids.insert(space_id);
|
|
|
|
}
|
2018-09-14 10:35:18 +02:00
|
|
|
|
2018-09-11 21:59:35 +02:00
|
|
|
/*
|
|
|
|
Retrieve default data directory, to be used with --copy-back.
|
|
|
|
|
|
|
|
On Windows, default datadir is ..\data, relative to the
|
|
|
|
directory where mariabackup.exe is located(usually "bin")
|
|
|
|
|
|
|
|
Elsewhere, the compiled-in constant MYSQL_DATADIR is used.
|
|
|
|
*/
|
|
|
|
static char *get_default_datadir() {
|
|
|
|
static char ddir[] = MYSQL_DATADIR;
|
|
|
|
#ifdef _WIN32
|
|
|
|
static char buf[MAX_PATH];
|
|
|
|
DWORD size = (DWORD)sizeof(buf) - 1;
|
|
|
|
if (GetModuleFileName(NULL, buf, size) <= size)
|
|
|
|
{
|
|
|
|
char *p;
|
|
|
|
if ((p = strrchr(buf, '\\')))
|
|
|
|
{
|
|
|
|
*p = 0;
|
|
|
|
if ((p = strrchr(buf, '\\')))
|
|
|
|
{
|
|
|
|
strncpy(p + 1, "data", buf + MAX_PATH - p);
|
|
|
|
return buf;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
return ddir;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* ======== Date copying thread context ======== */
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
datafiles_iter_t *it;
|
|
|
|
uint num;
|
|
|
|
uint *count;
|
2018-09-20 20:27:59 +02:00
|
|
|
pthread_mutex_t* count_mutex;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
CorruptedPages *corrupted_pages;
|
2017-04-18 20:43:20 +02:00
|
|
|
} data_thread_ctxt_t;
|
|
|
|
|
|
|
|
/* ======== for option and variables ======== */
|
2018-05-08 16:08:08 +02:00
|
|
|
#include <../../client/client_priv.h>
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
enum options_xtrabackup
|
|
|
|
{
|
2020-04-02 23:43:09 +02:00
|
|
|
OPT_XTRA_TARGET_DIR= 1000, /* make sure it is larger
|
|
|
|
than OPT_MAX_CLIENT_OPTION */
|
2017-04-18 20:43:20 +02:00
|
|
|
OPT_XTRA_BACKUP,
|
|
|
|
OPT_XTRA_PREPARE,
|
|
|
|
OPT_XTRA_EXPORT,
|
|
|
|
OPT_XTRA_PRINT_PARAM,
|
|
|
|
OPT_XTRA_USE_MEMORY,
|
|
|
|
OPT_XTRA_THROTTLE,
|
|
|
|
OPT_XTRA_LOG_COPY_INTERVAL,
|
|
|
|
OPT_XTRA_INCREMENTAL,
|
|
|
|
OPT_XTRA_INCREMENTAL_BASEDIR,
|
|
|
|
OPT_XTRA_EXTRA_LSNDIR,
|
|
|
|
OPT_XTRA_INCREMENTAL_DIR,
|
|
|
|
OPT_XTRA_TABLES,
|
|
|
|
OPT_XTRA_TABLES_FILE,
|
|
|
|
OPT_XTRA_DATABASES,
|
|
|
|
OPT_XTRA_DATABASES_FILE,
|
|
|
|
OPT_XTRA_PARALLEL,
|
2018-12-20 14:31:18 +01:00
|
|
|
OPT_XTRA_EXTENDED_VALIDATION,
|
2019-01-23 13:00:12 +01:00
|
|
|
OPT_XTRA_ENCRYPTED_BACKUP,
|
2017-04-18 20:43:20 +02:00
|
|
|
OPT_XTRA_STREAM,
|
|
|
|
OPT_XTRA_COMPRESS,
|
|
|
|
OPT_XTRA_COMPRESS_THREADS,
|
|
|
|
OPT_XTRA_COMPRESS_CHUNK_SIZE,
|
|
|
|
OPT_LOG,
|
|
|
|
OPT_INNODB,
|
|
|
|
OPT_INNODB_DATA_FILE_PATH,
|
|
|
|
OPT_INNODB_DATA_HOME_DIR,
|
|
|
|
OPT_INNODB_ADAPTIVE_HASH_INDEX,
|
|
|
|
OPT_INNODB_DOUBLEWRITE,
|
|
|
|
OPT_INNODB_FILE_PER_TABLE,
|
2018-04-30 17:12:14 +02:00
|
|
|
OPT_INNODB_FLUSH_METHOD,
|
2017-04-18 20:43:20 +02:00
|
|
|
OPT_INNODB_LOG_GROUP_HOME_DIR,
|
|
|
|
OPT_INNODB_MAX_DIRTY_PAGES_PCT,
|
|
|
|
OPT_INNODB_MAX_PURGE_LAG,
|
|
|
|
OPT_INNODB_STATUS_FILE,
|
|
|
|
OPT_INNODB_AUTOEXTEND_INCREMENT,
|
|
|
|
OPT_INNODB_BUFFER_POOL_SIZE,
|
|
|
|
OPT_INNODB_COMMIT_CONCURRENCY,
|
|
|
|
OPT_INNODB_CONCURRENCY_TICKETS,
|
|
|
|
OPT_INNODB_FILE_IO_THREADS,
|
|
|
|
OPT_INNODB_IO_CAPACITY,
|
|
|
|
OPT_INNODB_READ_IO_THREADS,
|
|
|
|
OPT_INNODB_WRITE_IO_THREADS,
|
|
|
|
OPT_INNODB_USE_NATIVE_AIO,
|
|
|
|
OPT_INNODB_PAGE_SIZE,
|
|
|
|
OPT_INNODB_BUFFER_POOL_FILENAME,
|
|
|
|
OPT_INNODB_LOCK_WAIT_TIMEOUT,
|
|
|
|
OPT_INNODB_LOG_BUFFER_SIZE,
|
2022-10-17 16:05:26 +02:00
|
|
|
#if defined __linux__ || defined _WIN32
|
|
|
|
OPT_INNODB_LOG_FILE_BUFFERING,
|
|
|
|
#endif
|
MDEV-30136: Deprecate innodb_flush_method
We introduce the following settable Boolean global variables:
innodb_log_file_write_through: Whether writes to ib_logfile0 are
write-through (disabling any caching, as in O_SYNC or O_DSYNC).
innodb_data_file_write_through: Whether writes to any InnoDB data files
(including the temporary tablespace) are write-through.
innodb_data_file_buffering: Whether the file system cache is enabled
for InnoDB data files.
All these parameters are OFF by default, that is, the file system cache
will be disabled, but any hardware caching is enabled, that is,
explicit calls to fsync(), fdatasync() or similar functions are needed.
On systems that support FUA it may make sense to enable write-through,
to avoid extra system calls.
If the deprecated read-only start-up parameter is set to one of the
following values, then the values of the 4 Boolean flags (the above 3
plus innodb_log_file_buffering) will be set as follows:
O_DSYNC:
innodb_log_file_write_through=ON, innodb_data_file_write_through=ON,
innodb_data_file_buffering=OFF, and
(if supported) innodb_log_file_buffering=OFF.
fsync, littlesync, nosync, or (Microsoft Windows specific) normal:
innodb_log_file_write_through=OFF, innodb_data_file_write_through=OFF,
and innodb_data_file_buffering=ON.
Note: fsync() or fdatasync() will only be disabled if the separate
parameter debug_no_sync (in the code, my_disable_sync) is set.
In mariadb-backup, the parameter innodb_flush_method will be ignored.
The Boolean parameters can be modified by SET GLOBAL while the
server is running. This will require reopening the ib_logfile0
or all currently open InnoDB data files.
We will open files straight in O_DSYNC or O_SYNC mode when applicable.
Data files we will try to open straight in O_DIRECT mode when the
page size is at least 4096 bytes. For atomically creating data files,
we will invoke os_file_set_nocache() to enable O_DIRECT afterwards,
because O_DIRECT is not supported on some file systems. We will also
continue to invoke os_file_set_nocache() on ib_logfile0 when
innodb_log_file_buffering=OFF can be fulfilled.
For reopening the ib_logfile0, we use the same logic that was developed
for online log resizing and reused for updates of
innodb_log_file_buffering.
Reopening all data files is implemented in the new function
fil_space_t::reopen_all().
Reviewed by: Vladislav Vaintroub
Tested by: Matthias Leich
2022-12-14 13:43:32 +01:00
|
|
|
OPT_INNODB_DATA_FILE_BUFFERING,
|
|
|
|
OPT_INNODB_DATA_FILE_WRITE_THROUGH,
|
2017-04-18 20:43:20 +02:00
|
|
|
OPT_INNODB_LOG_FILE_SIZE,
|
|
|
|
OPT_INNODB_OPEN_FILES,
|
|
|
|
OPT_XTRA_DEBUG_SYNC,
|
|
|
|
OPT_INNODB_CHECKSUM_ALGORITHM,
|
|
|
|
OPT_INNODB_UNDO_DIRECTORY,
|
|
|
|
OPT_INNODB_UNDO_TABLESPACES,
|
|
|
|
OPT_XTRA_INCREMENTAL_FORCE_SCAN,
|
|
|
|
OPT_DEFAULTS_GROUP,
|
|
|
|
OPT_CLOSE_FILES,
|
|
|
|
OPT_CORE_FILE,
|
|
|
|
|
|
|
|
OPT_COPY_BACK,
|
|
|
|
OPT_MOVE_BACK,
|
|
|
|
OPT_GALERA_INFO,
|
|
|
|
OPT_SLAVE_INFO,
|
|
|
|
OPT_NO_LOCK,
|
|
|
|
OPT_SAFE_SLAVE_BACKUP,
|
|
|
|
OPT_RSYNC,
|
|
|
|
OPT_FORCE_NON_EMPTY_DIRS,
|
|
|
|
OPT_NO_VERSION_CHECK,
|
|
|
|
OPT_NO_BACKUP_LOCKS,
|
|
|
|
OPT_DECOMPRESS,
|
|
|
|
OPT_INCREMENTAL_HISTORY_NAME,
|
|
|
|
OPT_INCREMENTAL_HISTORY_UUID,
|
|
|
|
OPT_REMOVE_ORIGINAL,
|
|
|
|
OPT_LOCK_WAIT_QUERY_TYPE,
|
|
|
|
OPT_KILL_LONG_QUERY_TYPE,
|
|
|
|
OPT_HISTORY,
|
|
|
|
OPT_KILL_LONG_QUERIES_TIMEOUT,
|
|
|
|
OPT_LOCK_WAIT_TIMEOUT,
|
|
|
|
OPT_LOCK_WAIT_THRESHOLD,
|
|
|
|
OPT_DEBUG_SLEEP_BEFORE_UNLOCK,
|
|
|
|
OPT_SAFE_SLAVE_BACKUP_TIMEOUT,
|
|
|
|
OPT_BINLOG_INFO,
|
|
|
|
OPT_XB_SECURE_AUTH,
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
OPT_XTRA_TABLES_EXCLUDE,
|
|
|
|
OPT_XTRA_DATABASES_EXCLUDE,
|
2017-09-11 18:45:36 +02:00
|
|
|
OPT_PROTOCOL,
|
2019-03-08 19:45:45 +01:00
|
|
|
OPT_INNODB_COMPRESSION_LEVEL,
|
2018-06-07 16:13:54 +02:00
|
|
|
OPT_LOCK_DDL_PER_TABLE,
|
|
|
|
OPT_ROCKSDB_DATADIR,
|
2018-12-14 23:36:21 +01:00
|
|
|
OPT_BACKUP_ROCKSDB,
|
2020-06-08 10:45:56 +02:00
|
|
|
OPT_XTRA_CHECK_PRIVILEGES,
|
2020-12-01 18:51:14 +01:00
|
|
|
OPT_XTRA_MYSQLD_ARGS,
|
2021-04-01 11:33:59 +02:00
|
|
|
OPT_XB_IGNORE_INNODB_PAGE_CORRUPTION,
|
2022-06-28 09:16:31 +02:00
|
|
|
OPT_INNODB_FORCE_RECOVERY
|
2017-04-18 20:43:20 +02:00
|
|
|
};
|
|
|
|
|
2020-04-02 23:43:09 +02:00
|
|
|
struct my_option xb_client_options[]= {
|
|
|
|
{"verbose", 'V', "display verbose output", (G_PTR *) &verbose,
|
|
|
|
(G_PTR *) &verbose, 0, GET_BOOL, NO_ARG, FALSE, 0, 0, 0, 0, 0},
|
2021-05-31 12:27:47 +02:00
|
|
|
{"version", 'v', "print version information",
|
2020-04-02 23:43:09 +02:00
|
|
|
(G_PTR *) &xtrabackup_version, (G_PTR *) &xtrabackup_version, 0, GET_BOOL,
|
|
|
|
NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"target-dir", OPT_XTRA_TARGET_DIR, "destination directory",
|
|
|
|
(G_PTR *) &xtrabackup_target_dir, (G_PTR *) &xtrabackup_target_dir, 0,
|
|
|
|
GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"backup", OPT_XTRA_BACKUP, "take backup to target-dir",
|
|
|
|
(G_PTR *) &xtrabackup_backup, (G_PTR *) &xtrabackup_backup, 0, GET_BOOL,
|
|
|
|
NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"prepare", OPT_XTRA_PREPARE,
|
|
|
|
"prepare a backup for starting mysql server on the backup.",
|
|
|
|
(G_PTR *) &xtrabackup_prepare, (G_PTR *) &xtrabackup_prepare, 0, GET_BOOL,
|
|
|
|
NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"export", OPT_XTRA_EXPORT,
|
|
|
|
"create files to import to another database when prepare.",
|
|
|
|
(G_PTR *) &xtrabackup_export, (G_PTR *) &xtrabackup_export, 0, GET_BOOL,
|
|
|
|
NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"print-param", OPT_XTRA_PRINT_PARAM,
|
|
|
|
"print parameter of mysqld needed for copyback.",
|
|
|
|
(G_PTR *) &xtrabackup_print_param, (G_PTR *) &xtrabackup_print_param, 0,
|
|
|
|
GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"use-memory", OPT_XTRA_USE_MEMORY,
|
|
|
|
"The value is used instead of buffer_pool_size",
|
|
|
|
(G_PTR *) &xtrabackup_use_memory, (G_PTR *) &xtrabackup_use_memory, 0,
|
|
|
|
GET_LL, REQUIRED_ARG, 100 * 1024 * 1024L, 1024 * 1024L, LONGLONG_MAX, 0,
|
|
|
|
1024 * 1024L, 0},
|
|
|
|
{"throttle", OPT_XTRA_THROTTLE,
|
|
|
|
"limit count of IO operations (pairs of read&write) per second to IOS "
|
|
|
|
"values (for '--backup')",
|
|
|
|
(G_PTR *) &xtrabackup_throttle, (G_PTR *) &xtrabackup_throttle, 0,
|
|
|
|
GET_LONG, REQUIRED_ARG, 0, 0, LONG_MAX, 0, 1, 0},
|
|
|
|
{"log", OPT_LOG, "Ignored option for MySQL option compatibility",
|
|
|
|
(G_PTR *) &log_ignored_opt, (G_PTR *) &log_ignored_opt, 0, GET_STR,
|
|
|
|
OPT_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"log-copy-interval", OPT_XTRA_LOG_COPY_INTERVAL,
|
|
|
|
"time interval between checks done by log copying thread in milliseconds "
|
|
|
|
"(default is 1 second).",
|
|
|
|
(G_PTR *) &xtrabackup_log_copy_interval,
|
|
|
|
(G_PTR *) &xtrabackup_log_copy_interval, 0, GET_LONG, REQUIRED_ARG, 1000,
|
|
|
|
0, LONG_MAX, 0, 1, 0},
|
|
|
|
{"extra-lsndir", OPT_XTRA_EXTRA_LSNDIR,
|
|
|
|
"(for --backup): save an extra copy of the xtrabackup_checkpoints file "
|
|
|
|
"in this directory.",
|
|
|
|
(G_PTR *) &xtrabackup_extra_lsndir, (G_PTR *) &xtrabackup_extra_lsndir, 0,
|
|
|
|
GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"incremental-lsn", OPT_XTRA_INCREMENTAL,
|
|
|
|
"(for --backup): copy only .ibd pages newer than specified LSN "
|
|
|
|
"'high:low'. ##ATTENTION##: If a wrong LSN value is specified, it is "
|
|
|
|
"impossible to diagnose this, causing the backup to be unusable. Be "
|
|
|
|
"careful!",
|
|
|
|
(G_PTR *) &xtrabackup_incremental, (G_PTR *) &xtrabackup_incremental, 0,
|
|
|
|
GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"incremental-basedir", OPT_XTRA_INCREMENTAL_BASEDIR,
|
|
|
|
"(for --backup): copy only .ibd pages newer than backup at specified "
|
|
|
|
"directory.",
|
|
|
|
(G_PTR *) &xtrabackup_incremental_basedir,
|
|
|
|
(G_PTR *) &xtrabackup_incremental_basedir, 0, GET_STR, REQUIRED_ARG, 0, 0,
|
|
|
|
0, 0, 0, 0},
|
|
|
|
{"incremental-dir", OPT_XTRA_INCREMENTAL_DIR,
|
|
|
|
"(for --prepare): apply .delta files and logfile in the specified "
|
|
|
|
"directory.",
|
|
|
|
(G_PTR *) &xtrabackup_incremental_dir,
|
|
|
|
(G_PTR *) &xtrabackup_incremental_dir, 0, GET_STR, REQUIRED_ARG, 0, 0, 0,
|
|
|
|
0, 0, 0},
|
|
|
|
{"tables", OPT_XTRA_TABLES, "filtering by regexp for table names.",
|
|
|
|
(G_PTR *) &xtrabackup_tables, (G_PTR *) &xtrabackup_tables, 0, GET_STR,
|
|
|
|
REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"tables_file", OPT_XTRA_TABLES_FILE,
|
|
|
|
"filtering by list of the exact database.table name in the file.",
|
|
|
|
(G_PTR *) &xtrabackup_tables_file, (G_PTR *) &xtrabackup_tables_file, 0,
|
|
|
|
GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"databases", OPT_XTRA_DATABASES, "filtering by list of databases.",
|
|
|
|
(G_PTR *) &xtrabackup_databases, (G_PTR *) &xtrabackup_databases, 0,
|
|
|
|
GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"databases_file", OPT_XTRA_DATABASES_FILE,
|
|
|
|
"filtering by list of databases in the file.",
|
|
|
|
(G_PTR *) &xtrabackup_databases_file,
|
|
|
|
(G_PTR *) &xtrabackup_databases_file, 0, GET_STR, REQUIRED_ARG, 0, 0, 0,
|
|
|
|
0, 0, 0},
|
|
|
|
{"tables-exclude", OPT_XTRA_TABLES_EXCLUDE,
|
|
|
|
"filtering by regexp for table names. "
|
|
|
|
"Operates the same way as --tables, but matched names are excluded from "
|
|
|
|
"backup. "
|
|
|
|
"Note that this option has a higher priority than --tables.",
|
|
|
|
(G_PTR *) &xtrabackup_tables_exclude,
|
|
|
|
(G_PTR *) &xtrabackup_tables_exclude, 0, GET_STR, REQUIRED_ARG, 0, 0, 0,
|
|
|
|
0, 0, 0},
|
|
|
|
{"databases-exclude", OPT_XTRA_DATABASES_EXCLUDE,
|
|
|
|
"Excluding databases based on name, "
|
|
|
|
"Operates the same way as --databases, but matched names are excluded "
|
|
|
|
"from backup. "
|
|
|
|
"Note that this option has a higher priority than --databases.",
|
|
|
|
(G_PTR *) &xtrabackup_databases_exclude,
|
|
|
|
(G_PTR *) &xtrabackup_databases_exclude, 0, GET_STR, REQUIRED_ARG, 0, 0,
|
|
|
|
0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"stream", OPT_XTRA_STREAM,
|
|
|
|
"Stream all backup files to the standard output "
|
|
|
|
"in the specified format."
|
|
|
|
"Supported format is 'mbstream' or 'xbstream'.",
|
|
|
|
(G_PTR *) &xtrabackup_stream_str, (G_PTR *) &xtrabackup_stream_str, 0,
|
|
|
|
GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"compress", OPT_XTRA_COMPRESS,
|
|
|
|
"Compress individual backup files using the "
|
|
|
|
"specified compression algorithm. Currently the only supported algorithm "
|
|
|
|
"is 'quicklz'. It is also the default algorithm, i.e. the one used when "
|
|
|
|
"--compress is used without an argument.",
|
|
|
|
(G_PTR *) &xtrabackup_compress_alg, (G_PTR *) &xtrabackup_compress_alg, 0,
|
|
|
|
GET_STR, OPT_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"compress-threads", OPT_XTRA_COMPRESS_THREADS,
|
|
|
|
"Number of threads for parallel data compression. The default value is "
|
|
|
|
"1.",
|
|
|
|
(G_PTR *) &xtrabackup_compress_threads,
|
|
|
|
(G_PTR *) &xtrabackup_compress_threads, 0, GET_UINT, REQUIRED_ARG, 1, 1,
|
|
|
|
UINT_MAX, 0, 0, 0},
|
|
|
|
|
|
|
|
{"compress-chunk-size", OPT_XTRA_COMPRESS_CHUNK_SIZE,
|
|
|
|
"Size of working buffer(s) for compression threads in bytes. The default "
|
|
|
|
"value is 64K.",
|
|
|
|
(G_PTR *) &xtrabackup_compress_chunk_size,
|
|
|
|
(G_PTR *) &xtrabackup_compress_chunk_size, 0, GET_ULL, REQUIRED_ARG,
|
|
|
|
(1 << 16), 1024, ULONGLONG_MAX, 0, 0, 0},
|
|
|
|
|
|
|
|
{"incremental-force-scan", OPT_XTRA_INCREMENTAL_FORCE_SCAN,
|
|
|
|
"Perform a full-scan incremental backup even in the presence of changed "
|
|
|
|
"page bitmap data",
|
|
|
|
(G_PTR *) &xtrabackup_incremental_force_scan,
|
|
|
|
(G_PTR *) &xtrabackup_incremental_force_scan, 0, GET_BOOL, NO_ARG, 0, 0,
|
|
|
|
0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"close_files", OPT_CLOSE_FILES,
|
|
|
|
"do not keep files opened. Use at your own "
|
|
|
|
"risk.",
|
|
|
|
(G_PTR *) &xb_close_files, (G_PTR *) &xb_close_files, 0, GET_BOOL, NO_ARG,
|
|
|
|
0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"core-file", OPT_CORE_FILE, "Write core on fatal signals", 0, 0, 0,
|
|
|
|
GET_NO_ARG, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"copy-back", OPT_COPY_BACK,
|
|
|
|
"Copy all the files in a previously made "
|
|
|
|
"backup from the backup directory to their original locations.",
|
|
|
|
(uchar *) &xtrabackup_copy_back, (uchar *) &xtrabackup_copy_back, 0,
|
|
|
|
GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"move-back", OPT_MOVE_BACK,
|
|
|
|
"Move all the files in a previously made "
|
|
|
|
"backup from the backup directory to the actual datadir location. "
|
|
|
|
"Use with caution, as it removes backup files.",
|
|
|
|
(uchar *) &xtrabackup_move_back, (uchar *) &xtrabackup_move_back, 0,
|
|
|
|
GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"galera-info", OPT_GALERA_INFO,
|
|
|
|
"This options creates the "
|
|
|
|
"xtrabackup_galera_info file which contains the local node state at "
|
|
|
|
"the time of the backup. Option should be used when performing the "
|
|
|
|
"backup of MariaDB Galera Cluster. Has no effect when backup locks "
|
|
|
|
"are used to create the backup.",
|
|
|
|
(uchar *) &opt_galera_info, (uchar *) &opt_galera_info, 0, GET_BOOL,
|
|
|
|
NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"slave-info", OPT_SLAVE_INFO,
|
|
|
|
"This option is useful when backing "
|
|
|
|
"up a replication slave server. It prints the binary log position "
|
|
|
|
"and name of the master server. It also writes this information to "
|
|
|
|
"the \"xtrabackup_slave_info\" file as a \"CHANGE MASTER\" command. "
|
|
|
|
"A new slave for this master can be set up by starting a slave server "
|
|
|
|
"on this backup and issuing a \"CHANGE MASTER\" command with the "
|
|
|
|
"binary log position saved in the \"xtrabackup_slave_info\" file.",
|
|
|
|
(uchar *) &opt_slave_info, (uchar *) &opt_slave_info, 0, GET_BOOL, NO_ARG,
|
|
|
|
0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"no-lock", OPT_NO_LOCK,
|
|
|
|
"Use this option to disable table lock "
|
|
|
|
"with \"FLUSH TABLES WITH READ LOCK\". Use it only if ALL your "
|
|
|
|
"tables are InnoDB and you DO NOT CARE about the binary log "
|
|
|
|
"position of the backup. This option shouldn't be used if there "
|
|
|
|
"are any DDL statements being executed or if any updates are "
|
|
|
|
"happening on non-InnoDB tables (this includes the system MyISAM "
|
|
|
|
"tables in the mysql database), otherwise it could lead to an "
|
|
|
|
"inconsistent backup. If you are considering to use --no-lock "
|
|
|
|
"because your backups are failing to acquire the lock, this could "
|
|
|
|
"be because of incoming replication events preventing the lock "
|
|
|
|
"from succeeding. Please try using --safe-slave-backup to "
|
|
|
|
"momentarily stop the replication slave thread, this may help "
|
|
|
|
"the backup to succeed and you then don't need to resort to "
|
|
|
|
"using this option.",
|
|
|
|
(uchar *) &opt_no_lock, (uchar *) &opt_no_lock, 0, GET_BOOL, NO_ARG, 0, 0,
|
|
|
|
0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"safe-slave-backup", OPT_SAFE_SLAVE_BACKUP,
|
|
|
|
"Stop slave SQL thread "
|
|
|
|
"and wait to start backup until Slave_open_temp_tables in "
|
|
|
|
"\"SHOW STATUS\" is zero. If there are no open temporary tables, "
|
|
|
|
"the backup will take place, otherwise the SQL thread will be "
|
|
|
|
"started and stopped until there are no open temporary tables. "
|
|
|
|
"The backup will fail if Slave_open_temp_tables does not become "
|
|
|
|
"zero after --safe-slave-backup-timeout seconds. The slave SQL "
|
|
|
|
"thread will be restarted when the backup finishes.",
|
|
|
|
(uchar *) &opt_safe_slave_backup, (uchar *) &opt_safe_slave_backup, 0,
|
|
|
|
GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"rsync", OPT_RSYNC,
|
|
|
|
"Uses the rsync utility to optimize local file "
|
|
|
|
"transfers. When this option is specified, innobackupex uses rsync "
|
|
|
|
"to copy all non-InnoDB files instead of spawning a separate cp for "
|
|
|
|
"each file, which can be much faster for servers with a large number "
|
|
|
|
"of databases or tables. This option cannot be used together with "
|
|
|
|
"--stream.",
|
|
|
|
(uchar *) &opt_rsync, (uchar *) &opt_rsync, 0, GET_BOOL, NO_ARG, 0, 0, 0,
|
|
|
|
0, 0, 0},
|
|
|
|
|
|
|
|
{"force-non-empty-directories", OPT_FORCE_NON_EMPTY_DIRS,
|
|
|
|
"This "
|
|
|
|
"option, when specified, makes --copy-back or --move-back transfer "
|
|
|
|
"files to non-empty directories. Note that no existing files will be "
|
2021-01-14 14:35:45 +01:00
|
|
|
"overwritten. If --copy-back or --move-back has to copy a file from "
|
2020-04-02 23:43:09 +02:00
|
|
|
"the backup directory which already exists in the destination "
|
|
|
|
"directory, it will still fail with an error.",
|
|
|
|
(uchar *) &opt_force_non_empty_dirs, (uchar *) &opt_force_non_empty_dirs,
|
|
|
|
0, GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"no-version-check", OPT_NO_VERSION_CHECK,
|
|
|
|
"This option disables the "
|
|
|
|
"version check which is enabled by the --version-check option.",
|
|
|
|
(uchar *) &opt_noversioncheck, (uchar *) &opt_noversioncheck, 0, GET_BOOL,
|
|
|
|
NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"no-backup-locks", OPT_NO_BACKUP_LOCKS,
|
|
|
|
"This option controls if "
|
|
|
|
"backup locks should be used instead of FLUSH TABLES WITH READ LOCK "
|
|
|
|
"on the backup stage. The option has no effect when backup locks are "
|
|
|
|
"not supported by the server. This option is enabled by default, "
|
|
|
|
"disable with --no-backup-locks.",
|
|
|
|
(uchar *) &opt_no_backup_locks, (uchar *) &opt_no_backup_locks, 0,
|
|
|
|
GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"decompress", OPT_DECOMPRESS,
|
|
|
|
"Decompresses all files with the .qp "
|
|
|
|
"extension in a backup previously made with the --compress option.",
|
|
|
|
(uchar *) &opt_decompress, (uchar *) &opt_decompress, 0, GET_BOOL, NO_ARG,
|
|
|
|
0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"user", 'u',
|
|
|
|
"This option specifies the MySQL username used "
|
|
|
|
"when connecting to the server, if that's not the current user. "
|
|
|
|
"The option accepts a string argument. See mysql --help for details.",
|
|
|
|
(uchar *) &opt_user, (uchar *) &opt_user, 0, GET_STR, REQUIRED_ARG, 0, 0,
|
|
|
|
0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"host", 'H',
|
|
|
|
"This option specifies the host to use when "
|
|
|
|
"connecting to the database server with TCP/IP. The option accepts "
|
|
|
|
"a string argument. See mysql --help for details.",
|
|
|
|
(uchar *) &opt_host, (uchar *) &opt_host, 0, GET_STR, REQUIRED_ARG, 0, 0,
|
|
|
|
0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"port", 'P',
|
|
|
|
"This option specifies the port to use when "
|
|
|
|
"connecting to the database server with TCP/IP. The option accepts "
|
|
|
|
"a string argument. See mysql --help for details.",
|
|
|
|
&opt_port, &opt_port, 0, GET_UINT, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"password", 'p',
|
|
|
|
"This option specifies the password to use "
|
|
|
|
"when connecting to the database. It accepts a string argument. "
|
|
|
|
"See mysql --help for details.",
|
|
|
|
0, 0, 0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"protocol", OPT_PROTOCOL,
|
|
|
|
"The protocol to use for connection (tcp, socket, pipe, memory).", 0, 0,
|
|
|
|
0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"socket", 'S',
|
|
|
|
"This option specifies the socket to use when "
|
|
|
|
"connecting to the local database server with a UNIX domain socket. "
|
|
|
|
"The option accepts a string argument. See mysql --help for details.",
|
|
|
|
(uchar *) &opt_socket, (uchar *) &opt_socket, 0, GET_STR, REQUIRED_ARG, 0,
|
|
|
|
0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"incremental-history-name", OPT_INCREMENTAL_HISTORY_NAME,
|
|
|
|
"This option specifies the name of the backup series stored in the "
|
2022-09-09 10:02:16 +02:00
|
|
|
XB_HISTORY_TABLE " history record to base an "
|
2020-04-02 23:43:09 +02:00
|
|
|
"incremental backup on. Xtrabackup will search the history table "
|
|
|
|
"looking for the most recent (highest innodb_to_lsn), successful "
|
|
|
|
"backup in the series and take the to_lsn value to use as the "
|
|
|
|
"starting lsn for the incremental backup. This will be mutually "
|
|
|
|
"exclusive with --incremental-history-uuid, --incremental-basedir "
|
|
|
|
"and --incremental-lsn. If no valid lsn can be found (no series by "
|
2021-05-31 12:27:47 +02:00
|
|
|
"that name, no successful backups by that name), an error will be returned."
|
|
|
|
" It is used with the --incremental option.",
|
2020-04-02 23:43:09 +02:00
|
|
|
(uchar *) &opt_incremental_history_name,
|
|
|
|
(uchar *) &opt_incremental_history_name, 0, GET_STR, REQUIRED_ARG, 0, 0,
|
|
|
|
0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"incremental-history-uuid", OPT_INCREMENTAL_HISTORY_UUID,
|
|
|
|
"This option specifies the UUID of the specific history record "
|
2022-09-09 10:02:16 +02:00
|
|
|
"stored in the " XB_HISTORY_TABLE " table to base an "
|
2020-04-02 23:43:09 +02:00
|
|
|
"incremental backup on. --incremental-history-name, "
|
|
|
|
"--incremental-basedir and --incremental-lsn. If no valid lsn can be "
|
2021-05-31 12:27:47 +02:00
|
|
|
"found (no success record with that uuid), an error will be returned."
|
|
|
|
" It is used with the --incremental option.",
|
2020-04-02 23:43:09 +02:00
|
|
|
(uchar *) &opt_incremental_history_uuid,
|
|
|
|
(uchar *) &opt_incremental_history_uuid, 0, GET_STR, REQUIRED_ARG, 0, 0,
|
|
|
|
0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"remove-original", OPT_REMOVE_ORIGINAL,
|
|
|
|
"Remove .qp files after decompression.", (uchar *) &opt_remove_original,
|
|
|
|
(uchar *) &opt_remove_original, 0, GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"ftwrl-wait-query-type", OPT_LOCK_WAIT_QUERY_TYPE,
|
|
|
|
"This option specifies which types of queries are allowed to complete "
|
|
|
|
"before innobackupex will issue the global lock. Default is all.",
|
|
|
|
(uchar *) &opt_lock_wait_query_type, (uchar *) &opt_lock_wait_query_type,
|
|
|
|
&query_type_typelib, GET_ENUM, REQUIRED_ARG, QUERY_TYPE_ALL, 0, 0, 0, 0,
|
|
|
|
0},
|
|
|
|
|
|
|
|
{"kill-long-query-type", OPT_KILL_LONG_QUERY_TYPE,
|
|
|
|
"This option specifies which types of queries should be killed to "
|
|
|
|
"unblock the global lock. Default is \"all\".",
|
|
|
|
(uchar *) &opt_kill_long_query_type, (uchar *) &opt_kill_long_query_type,
|
|
|
|
&query_type_typelib, GET_ENUM, REQUIRED_ARG, QUERY_TYPE_SELECT, 0, 0, 0,
|
|
|
|
0, 0},
|
|
|
|
|
|
|
|
{"history", OPT_HISTORY,
|
|
|
|
"This option enables the tracking of backup history in the "
|
2022-09-09 10:02:16 +02:00
|
|
|
XB_HISTORY_TABLE " table. An optional history "
|
2020-04-02 23:43:09 +02:00
|
|
|
"series name may be specified that will be placed with the history "
|
|
|
|
"record for the current backup being taken.",
|
|
|
|
NULL, NULL, 0, GET_STR, OPT_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"kill-long-queries-timeout", OPT_KILL_LONG_QUERIES_TIMEOUT,
|
|
|
|
"This option specifies the number of seconds innobackupex waits "
|
|
|
|
"between starting FLUSH TABLES WITH READ LOCK and killing those "
|
|
|
|
"queries that block it. Default is 0 seconds, which means "
|
|
|
|
"innobackupex will not attempt to kill any queries.",
|
|
|
|
(uchar *) &opt_kill_long_queries_timeout,
|
|
|
|
(uchar *) &opt_kill_long_queries_timeout, 0, GET_UINT, REQUIRED_ARG, 0, 0,
|
|
|
|
0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"ftwrl-wait-timeout", OPT_LOCK_WAIT_TIMEOUT,
|
|
|
|
"This option specifies time in seconds that innobackupex should wait "
|
|
|
|
"for queries that would block FTWRL before running it. If there are "
|
|
|
|
"still such queries when the timeout expires, innobackupex terminates "
|
|
|
|
"with an error. Default is 0, in which case innobackupex does not "
|
|
|
|
"wait for queries to complete and starts FTWRL immediately.",
|
|
|
|
(uchar *) &opt_lock_wait_timeout, (uchar *) &opt_lock_wait_timeout, 0,
|
|
|
|
GET_UINT, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"ftwrl-wait-threshold", OPT_LOCK_WAIT_THRESHOLD,
|
|
|
|
"This option specifies the query run time threshold which is used by "
|
|
|
|
"innobackupex to detect long-running queries with a non-zero value "
|
|
|
|
"of --ftwrl-wait-timeout. FTWRL is not started until such "
|
|
|
|
"long-running queries exist. This option has no effect if "
|
|
|
|
"--ftwrl-wait-timeout is 0. Default value is 60 seconds.",
|
|
|
|
(uchar *) &opt_lock_wait_threshold, (uchar *) &opt_lock_wait_threshold, 0,
|
|
|
|
GET_UINT, REQUIRED_ARG, 60, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
|
|
|
|
{"safe-slave-backup-timeout", OPT_SAFE_SLAVE_BACKUP_TIMEOUT,
|
|
|
|
"How many seconds --safe-slave-backup should wait for "
|
|
|
|
"Slave_open_temp_tables to become zero. (default 300)",
|
|
|
|
(uchar *) &opt_safe_slave_backup_timeout,
|
|
|
|
(uchar *) &opt_safe_slave_backup_timeout, 0, GET_UINT, REQUIRED_ARG, 300,
|
|
|
|
0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"binlog-info", OPT_BINLOG_INFO,
|
2021-05-31 12:27:47 +02:00
|
|
|
"This option controls how backup should retrieve server's binary log "
|
2020-04-02 23:43:09 +02:00
|
|
|
"coordinates corresponding to the backup. Possible values are OFF, ON, "
|
2021-05-31 12:27:47 +02:00
|
|
|
"LOCKLESS and AUTO.",
|
2020-04-02 23:43:09 +02:00
|
|
|
&opt_binlog_info, &opt_binlog_info, &binlog_info_typelib, GET_ENUM,
|
|
|
|
OPT_ARG, BINLOG_INFO_AUTO, 0, 0, 0, 0, 0},
|
|
|
|
|
|
|
|
{"secure-auth", OPT_XB_SECURE_AUTH,
|
|
|
|
"Refuse client connecting to server if it"
|
|
|
|
" uses old (pre-4.1.1) protocol.",
|
|
|
|
&opt_secure_auth, &opt_secure_auth, 0, GET_BOOL, NO_ARG, 1, 0, 0, 0, 0,
|
|
|
|
0},
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
|
|
|
|
{"log-innodb-page-corruption", OPT_XB_IGNORE_INNODB_PAGE_CORRUPTION,
|
|
|
|
"Continue backup if innodb corrupted pages are found. The pages are "
|
|
|
|
"logged in " MB_CORRUPTED_PAGES_FILE
|
|
|
|
" and backup is finished with error. "
|
|
|
|
"--prepare will try to fix corrupted pages. If " MB_CORRUPTED_PAGES_FILE
|
|
|
|
" exists after --prepare in base backup directory, backup still contains "
|
|
|
|
"corrupted pages and can not be considered as consistent.",
|
|
|
|
&opt_log_innodb_page_corruption, &opt_log_innodb_page_corruption, 0,
|
|
|
|
GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
2018-05-08 16:08:08 +02:00
|
|
|
#define MYSQL_CLIENT
|
2017-04-18 20:43:20 +02:00
|
|
|
#include "sslopt-longopts.h"
|
2018-05-08 16:08:08 +02:00
|
|
|
#undef MYSQL_CLIENT
|
2020-04-02 23:43:09 +02:00
|
|
|
{0, 0, 0, 0, 0, 0, GET_NO_ARG, NO_ARG, 0, 0, 0, 0, 0, 0}};
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
uint xb_client_options_count = array_elements(xb_client_options);
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#ifndef DBUG_OFF
|
|
|
|
/** Parameters to DBUG */
|
|
|
|
static const char *dbug_option;
|
|
|
|
#endif
|
|
|
|
|
2021-11-02 14:24:20 +01:00
|
|
|
#ifdef HAVE_URING
|
|
|
|
extern const char *io_uring_may_be_unsafe;
|
|
|
|
bool innodb_use_native_aio_default();
|
|
|
|
#endif
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
struct my_option xb_server_options[] =
|
|
|
|
{
|
|
|
|
{"datadir", 'h', "Path to the database root.", (G_PTR*) &mysql_data_home,
|
|
|
|
(G_PTR*) &mysql_data_home, 0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"tmpdir", 't',
|
|
|
|
"Path for temporary files. Several paths may be specified, separated by a "
|
2021-06-06 13:21:03 +02:00
|
|
|
#if defined(_WIN32)
|
2017-04-18 20:43:20 +02:00
|
|
|
"semicolon (;)"
|
|
|
|
#else
|
|
|
|
"colon (:)"
|
|
|
|
#endif
|
|
|
|
", in this case they are used in a round-robin fashion.",
|
|
|
|
(G_PTR*) &opt_mysql_tmpdir,
|
|
|
|
(G_PTR*) &opt_mysql_tmpdir, 0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
{"parallel", OPT_XTRA_PARALLEL,
|
2017-04-19 15:09:03 +02:00
|
|
|
"Number of threads to use for parallel datafiles transfer. "
|
|
|
|
"The default value is 1.",
|
2017-04-18 20:43:20 +02:00
|
|
|
(G_PTR*) &xtrabackup_parallel, (G_PTR*) &xtrabackup_parallel, 0, GET_INT,
|
|
|
|
REQUIRED_ARG, 1, 1, INT_MAX, 0, 0, 0},
|
|
|
|
|
2018-12-20 14:31:18 +01:00
|
|
|
{"extended_validation", OPT_XTRA_EXTENDED_VALIDATION,
|
MDEV-18129 Backup fails for encrypted tables: mariabackup: Database page corruption detected at page 1
If an encrypted table is created during backup, then
mariabackup --backup could wrongly fail.
This caused a failure of the test mariabackup.huge_lsn once on buildbot.
This is due to the way how InnoDB creates .ibd files. It would first
write a dummy page 0 with no encryption information. Due to this,
xb_fil_cur_open() could wrongly interpret that the table is not encrypted.
Subsequently, page_is_corrupted() would compare the computed page
checksum to the wrong checksum. (There are both "before" and "after"
checksums for encrypted pages.)
To work around this problem, we introduce a Boolean option
--backup-encrypted that is enabled by default. With this option,
Mariabackup will assume that a nonzero key_version implies that the
page is encrypted. We need this option in order to be able to copy
encrypted tables from MariaDB 10.1 or 10.2, because unencrypted pages
that were originally created before MySQL 5.1.48 could contain nonzero
garbage in the fields that were repurposed for encryption.
Later, MDEV-18128 would clean up the way how .ibd files are created,
to remove the need for this option.
page_is_corrupted(): Add missing const qualifiers, and do not check
space->crypt_data unless --skip-backup-encrypted has been specified.
xb_fil_cur_read(): After a failed page read, output a page dump.
2019-01-03 15:24:22 +01:00
|
|
|
"Enable extended validation for Innodb data pages during backup phase. "
|
|
|
|
"Will slow down backup considerably, in case encryption is used. "
|
|
|
|
"May fail if tables are created during the backup.",
|
2018-12-20 14:31:18 +01:00
|
|
|
(G_PTR*)&opt_extended_validation,
|
|
|
|
(G_PTR*)&opt_extended_validation,
|
|
|
|
0, GET_BOOL, NO_ARG, FALSE, 0, 0, 0, 0, 0},
|
|
|
|
|
2019-01-23 13:00:12 +01:00
|
|
|
{"encrypted_backup", OPT_XTRA_ENCRYPTED_BACKUP,
|
MDEV-18129 Backup fails for encrypted tables: mariabackup: Database page corruption detected at page 1
If an encrypted table is created during backup, then
mariabackup --backup could wrongly fail.
This caused a failure of the test mariabackup.huge_lsn once on buildbot.
This is due to the way how InnoDB creates .ibd files. It would first
write a dummy page 0 with no encryption information. Due to this,
xb_fil_cur_open() could wrongly interpret that the table is not encrypted.
Subsequently, page_is_corrupted() would compare the computed page
checksum to the wrong checksum. (There are both "before" and "after"
checksums for encrypted pages.)
To work around this problem, we introduce a Boolean option
--backup-encrypted that is enabled by default. With this option,
Mariabackup will assume that a nonzero key_version implies that the
page is encrypted. We need this option in order to be able to copy
encrypted tables from MariaDB 10.1 or 10.2, because unencrypted pages
that were originally created before MySQL 5.1.48 could contain nonzero
garbage in the fields that were repurposed for encryption.
Later, MDEV-18128 would clean up the way how .ibd files are created,
to remove the need for this option.
page_is_corrupted(): Add missing const qualifiers, and do not check
space->crypt_data unless --skip-backup-encrypted has been specified.
xb_fil_cur_read(): After a failed page read, output a page dump.
2019-01-03 15:24:22 +01:00
|
|
|
"In --backup, assume that nonzero key_version implies that the page"
|
2019-01-23 13:00:12 +01:00
|
|
|
" is encrypted. Use --backup --skip-encrypted-backup to allow"
|
MDEV-18129 Backup fails for encrypted tables: mariabackup: Database page corruption detected at page 1
If an encrypted table is created during backup, then
mariabackup --backup could wrongly fail.
This caused a failure of the test mariabackup.huge_lsn once on buildbot.
This is due to the way how InnoDB creates .ibd files. It would first
write a dummy page 0 with no encryption information. Due to this,
xb_fil_cur_open() could wrongly interpret that the table is not encrypted.
Subsequently, page_is_corrupted() would compare the computed page
checksum to the wrong checksum. (There are both "before" and "after"
checksums for encrypted pages.)
To work around this problem, we introduce a Boolean option
--backup-encrypted that is enabled by default. With this option,
Mariabackup will assume that a nonzero key_version implies that the
page is encrypted. We need this option in order to be able to copy
encrypted tables from MariaDB 10.1 or 10.2, because unencrypted pages
that were originally created before MySQL 5.1.48 could contain nonzero
garbage in the fields that were repurposed for encryption.
Later, MDEV-18128 would clean up the way how .ibd files are created,
to remove the need for this option.
page_is_corrupted(): Add missing const qualifiers, and do not check
space->crypt_data unless --skip-backup-encrypted has been specified.
xb_fil_cur_read(): After a failed page read, output a page dump.
2019-01-03 15:24:22 +01:00
|
|
|
" copying unencrypted that were originally created before MySQL 5.1.48.",
|
2019-01-23 13:00:12 +01:00
|
|
|
(G_PTR*)&opt_encrypted_backup,
|
|
|
|
(G_PTR*)&opt_encrypted_backup,
|
MDEV-18129 Backup fails for encrypted tables: mariabackup: Database page corruption detected at page 1
If an encrypted table is created during backup, then
mariabackup --backup could wrongly fail.
This caused a failure of the test mariabackup.huge_lsn once on buildbot.
This is due to the way how InnoDB creates .ibd files. It would first
write a dummy page 0 with no encryption information. Due to this,
xb_fil_cur_open() could wrongly interpret that the table is not encrypted.
Subsequently, page_is_corrupted() would compare the computed page
checksum to the wrong checksum. (There are both "before" and "after"
checksums for encrypted pages.)
To work around this problem, we introduce a Boolean option
--backup-encrypted that is enabled by default. With this option,
Mariabackup will assume that a nonzero key_version implies that the
page is encrypted. We need this option in order to be able to copy
encrypted tables from MariaDB 10.1 or 10.2, because unencrypted pages
that were originally created before MySQL 5.1.48 could contain nonzero
garbage in the fields that were repurposed for encryption.
Later, MDEV-18128 would clean up the way how .ibd files are created,
to remove the need for this option.
page_is_corrupted(): Add missing const qualifiers, and do not check
space->crypt_data unless --skip-backup-encrypted has been specified.
xb_fil_cur_read(): After a failed page read, output a page dump.
2019-01-03 15:24:22 +01:00
|
|
|
0, GET_BOOL, NO_ARG, TRUE, 0, 0, 0, 0, 0},
|
2018-12-20 14:31:18 +01:00
|
|
|
|
2022-01-18 19:07:01 +01:00
|
|
|
{"log", OPT_LOG, "Ignored option for MySQL option compatibility",
|
2017-04-18 20:43:20 +02:00
|
|
|
(G_PTR*) &log_ignored_opt, (G_PTR*) &log_ignored_opt, 0,
|
|
|
|
GET_STR, OPT_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
2022-01-18 19:07:01 +01:00
|
|
|
{"log_bin", OPT_LOG, "Base name for the log sequence",
|
2017-07-06 17:31:28 +02:00
|
|
|
&opt_log_bin, &opt_log_bin, 0, GET_STR, OPT_ARG, 0, 0, 0, 0, 0, 0},
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2022-01-18 19:07:01 +01:00
|
|
|
{"innodb", OPT_INNODB, "Ignored option for MySQL option compatibility",
|
2017-04-18 20:43:20 +02:00
|
|
|
(G_PTR*) &innobase_ignored_opt, (G_PTR*) &innobase_ignored_opt, 0,
|
|
|
|
GET_STR, OPT_ARG, 0, 0, 0, 0, 0, 0},
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#ifdef BTR_CUR_HASH_ADAPT
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_adaptive_hash_index", OPT_INNODB_ADAPTIVE_HASH_INDEX,
|
MDEV-20487 Set innodb_adaptive_hash_index=OFF by default
Based on the performance testing that was conducted in MDEV-17492,
the InnoDB adaptive hash index could only help performance in specific,
almost-read-only workloads. It could slow down all kinds of workloads
(especially DROP TABLE, TRUNCATE TABLE, ALTER TABLE, or DROP INDEX
operations), and it can become corrupted, causing crashes (such as
MDEV-18815, MDEV-20203) and possibly data corruption. Furthermore,
the adaptive hash index consumes space from the InnoDB buffer pool,
which could hurt performance when the working set would almost fit
in the buffer pool.
Given all this, it is best to disable the adaptive hash index by default.
2019-10-23 16:43:31 +02:00
|
|
|
"Enable InnoDB adaptive hash index (disabled by default).",
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
&btr_search_enabled,
|
|
|
|
&btr_search_enabled,
|
MDEV-20487 Set innodb_adaptive_hash_index=OFF by default
Based on the performance testing that was conducted in MDEV-17492,
the InnoDB adaptive hash index could only help performance in specific,
almost-read-only workloads. It could slow down all kinds of workloads
(especially DROP TABLE, TRUNCATE TABLE, ALTER TABLE, or DROP INDEX
operations), and it can become corrupted, causing crashes (such as
MDEV-18815, MDEV-20203) and possibly data corruption. Furthermore,
the adaptive hash index consumes space from the InnoDB buffer pool,
which could hurt performance when the working set would almost fit
in the buffer pool.
Given all this, it is best to disable the adaptive hash index by default.
2019-10-23 16:43:31 +02:00
|
|
|
0, GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#endif /* BTR_CUR_HASH_ADAPT */
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_autoextend_increment", OPT_INNODB_AUTOEXTEND_INCREMENT,
|
|
|
|
"Data file autoextend increment in megabytes",
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
(G_PTR*) &sys_tablespace_auto_extend_increment,
|
|
|
|
(G_PTR*) &sys_tablespace_auto_extend_increment,
|
2020-10-16 06:51:37 +02:00
|
|
|
0, GET_UINT, REQUIRED_ARG, 8, 1, 1000, 0, 1, 0},
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_data_file_path", OPT_INNODB_DATA_FILE_PATH,
|
|
|
|
"Path to individual files and their sizes.", &innobase_data_file_path,
|
2017-07-06 17:31:28 +02:00
|
|
|
&innobase_data_file_path, 0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_data_home_dir", OPT_INNODB_DATA_HOME_DIR,
|
|
|
|
"The common part for InnoDB table spaces.", &innobase_data_home_dir,
|
2017-07-06 17:31:28 +02:00
|
|
|
&innobase_data_home_dir, 0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
{"innodb_doublewrite", OPT_INNODB_DOUBLEWRITE,
|
|
|
|
"Enable InnoDB doublewrite buffer during --prepare.",
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
(G_PTR*) &srv_use_doublewrite_buf,
|
|
|
|
(G_PTR*) &srv_use_doublewrite_buf, 0, GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_io_capacity", OPT_INNODB_IO_CAPACITY,
|
|
|
|
"Number of IOPs the server can do. Tunes the background IO rate",
|
|
|
|
(G_PTR*) &srv_io_capacity, (G_PTR*) &srv_io_capacity,
|
|
|
|
0, GET_ULONG, OPT_ARG, 200, 100, ~0UL, 0, 0, 0},
|
|
|
|
{"innodb_file_io_threads", OPT_INNODB_FILE_IO_THREADS,
|
|
|
|
"Number of file I/O threads in InnoDB.", (G_PTR*) &innobase_file_io_threads,
|
|
|
|
(G_PTR*) &innobase_file_io_threads, 0, GET_LONG, REQUIRED_ARG, 4, 4, 64, 0,
|
|
|
|
1, 0},
|
|
|
|
{"innodb_read_io_threads", OPT_INNODB_READ_IO_THREADS,
|
|
|
|
"Number of background read I/O threads in InnoDB.", (G_PTR*) &innobase_read_io_threads,
|
|
|
|
(G_PTR*) &innobase_read_io_threads, 0, GET_LONG, REQUIRED_ARG, 4, 1, 64, 0,
|
|
|
|
1, 0},
|
|
|
|
{"innodb_write_io_threads", OPT_INNODB_WRITE_IO_THREADS,
|
|
|
|
"Number of background write I/O threads in InnoDB.", (G_PTR*) &innobase_write_io_threads,
|
|
|
|
(G_PTR*) &innobase_write_io_threads, 0, GET_LONG, REQUIRED_ARG, 4, 1, 64, 0,
|
|
|
|
1, 0},
|
|
|
|
{"innodb_file_per_table", OPT_INNODB_FILE_PER_TABLE,
|
|
|
|
"Stores each InnoDB table to an .ibd file in the database dir.",
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
(G_PTR*) &srv_file_per_table,
|
|
|
|
(G_PTR*) &srv_file_per_table, 0, GET_BOOL, NO_ARG,
|
2017-04-18 20:43:20 +02:00
|
|
|
FALSE, 0, 0, 0, 0, 0},
|
2017-04-18 21:05:57 +02:00
|
|
|
|
2018-04-30 17:12:14 +02:00
|
|
|
{"innodb_flush_method", OPT_INNODB_FLUSH_METHOD,
|
MDEV-30136: Deprecate innodb_flush_method
We introduce the following settable Boolean global variables:
innodb_log_file_write_through: Whether writes to ib_logfile0 are
write-through (disabling any caching, as in O_SYNC or O_DSYNC).
innodb_data_file_write_through: Whether writes to any InnoDB data files
(including the temporary tablespace) are write-through.
innodb_data_file_buffering: Whether the file system cache is enabled
for InnoDB data files.
All these parameters are OFF by default, that is, the file system cache
will be disabled, but any hardware caching is enabled, that is,
explicit calls to fsync(), fdatasync() or similar functions are needed.
On systems that support FUA it may make sense to enable write-through,
to avoid extra system calls.
If the deprecated read-only start-up parameter is set to one of the
following values, then the values of the 4 Boolean flags (the above 3
plus innodb_log_file_buffering) will be set as follows:
O_DSYNC:
innodb_log_file_write_through=ON, innodb_data_file_write_through=ON,
innodb_data_file_buffering=OFF, and
(if supported) innodb_log_file_buffering=OFF.
fsync, littlesync, nosync, or (Microsoft Windows specific) normal:
innodb_log_file_write_through=OFF, innodb_data_file_write_through=OFF,
and innodb_data_file_buffering=ON.
Note: fsync() or fdatasync() will only be disabled if the separate
parameter debug_no_sync (in the code, my_disable_sync) is set.
In mariadb-backup, the parameter innodb_flush_method will be ignored.
The Boolean parameters can be modified by SET GLOBAL while the
server is running. This will require reopening the ib_logfile0
or all currently open InnoDB data files.
We will open files straight in O_DSYNC or O_SYNC mode when applicable.
Data files we will try to open straight in O_DIRECT mode when the
page size is at least 4096 bytes. For atomically creating data files,
we will invoke os_file_set_nocache() to enable O_DIRECT afterwards,
because O_DIRECT is not supported on some file systems. We will also
continue to invoke os_file_set_nocache() on ib_logfile0 when
innodb_log_file_buffering=OFF can be fulfilled.
For reopening the ib_logfile0, we use the same logic that was developed
for online log resizing and reused for updates of
innodb_log_file_buffering.
Reopening all data files is implemented in the new function
fil_space_t::reopen_all().
Reviewed by: Vladislav Vaintroub
Tested by: Matthias Leich
2022-12-14 13:43:32 +01:00
|
|
|
"Ignored parameter with no effect",
|
|
|
|
&innodb_flush_method, &innodb_flush_method,
|
2018-04-30 17:12:14 +02:00
|
|
|
&innodb_flush_method_typelib, GET_ENUM, REQUIRED_ARG,
|
MDEV-30136: Deprecate innodb_flush_method
We introduce the following settable Boolean global variables:
innodb_log_file_write_through: Whether writes to ib_logfile0 are
write-through (disabling any caching, as in O_SYNC or O_DSYNC).
innodb_data_file_write_through: Whether writes to any InnoDB data files
(including the temporary tablespace) are write-through.
innodb_data_file_buffering: Whether the file system cache is enabled
for InnoDB data files.
All these parameters are OFF by default, that is, the file system cache
will be disabled, but any hardware caching is enabled, that is,
explicit calls to fsync(), fdatasync() or similar functions are needed.
On systems that support FUA it may make sense to enable write-through,
to avoid extra system calls.
If the deprecated read-only start-up parameter is set to one of the
following values, then the values of the 4 Boolean flags (the above 3
plus innodb_log_file_buffering) will be set as follows:
O_DSYNC:
innodb_log_file_write_through=ON, innodb_data_file_write_through=ON,
innodb_data_file_buffering=OFF, and
(if supported) innodb_log_file_buffering=OFF.
fsync, littlesync, nosync, or (Microsoft Windows specific) normal:
innodb_log_file_write_through=OFF, innodb_data_file_write_through=OFF,
and innodb_data_file_buffering=ON.
Note: fsync() or fdatasync() will only be disabled if the separate
parameter debug_no_sync (in the code, my_disable_sync) is set.
In mariadb-backup, the parameter innodb_flush_method will be ignored.
The Boolean parameters can be modified by SET GLOBAL while the
server is running. This will require reopening the ib_logfile0
or all currently open InnoDB data files.
We will open files straight in O_DSYNC or O_SYNC mode when applicable.
Data files we will try to open straight in O_DIRECT mode when the
page size is at least 4096 bytes. For atomically creating data files,
we will invoke os_file_set_nocache() to enable O_DIRECT afterwards,
because O_DIRECT is not supported on some file systems. We will also
continue to invoke os_file_set_nocache() on ib_logfile0 when
innodb_log_file_buffering=OFF can be fulfilled.
For reopening the ib_logfile0, we use the same logic that was developed
for online log resizing and reused for updates of
innodb_log_file_buffering.
Reopening all data files is implemented in the new function
fil_space_t::reopen_all().
Reviewed by: Vladislav Vaintroub
Tested by: Matthias Leich
2022-12-14 13:43:32 +01:00
|
|
|
4/* O_DIRECT */, 0, 0, 0, 0, 0},
|
2018-04-30 17:12:14 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_log_buffer_size", OPT_INNODB_LOG_BUFFER_SIZE,
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
"Redo log buffer size in bytes.",
|
|
|
|
(G_PTR*) &log_sys.buf_size, (G_PTR*) &log_sys.buf_size, 0,
|
|
|
|
IF_WIN(GET_ULL,GET_ULONG), REQUIRED_ARG, 2U << 20,
|
|
|
|
2U << 20, SIZE_T_MAX, 0, 4096, 0},
|
2022-10-17 16:05:26 +02:00
|
|
|
#if defined __linux__ || defined _WIN32
|
|
|
|
{"innodb_log_file_buffering", OPT_INNODB_LOG_FILE_BUFFERING,
|
|
|
|
"Whether the file system cache for ib_logfile0 is enabled during --backup",
|
|
|
|
(G_PTR*) &log_sys.log_buffered,
|
|
|
|
(G_PTR*) &log_sys.log_buffered, 0, GET_BOOL, NO_ARG,
|
|
|
|
TRUE, 0, 0, 0, 0, 0},
|
|
|
|
#endif
|
MDEV-30136: Deprecate innodb_flush_method
We introduce the following settable Boolean global variables:
innodb_log_file_write_through: Whether writes to ib_logfile0 are
write-through (disabling any caching, as in O_SYNC or O_DSYNC).
innodb_data_file_write_through: Whether writes to any InnoDB data files
(including the temporary tablespace) are write-through.
innodb_data_file_buffering: Whether the file system cache is enabled
for InnoDB data files.
All these parameters are OFF by default, that is, the file system cache
will be disabled, but any hardware caching is enabled, that is,
explicit calls to fsync(), fdatasync() or similar functions are needed.
On systems that support FUA it may make sense to enable write-through,
to avoid extra system calls.
If the deprecated read-only start-up parameter is set to one of the
following values, then the values of the 4 Boolean flags (the above 3
plus innodb_log_file_buffering) will be set as follows:
O_DSYNC:
innodb_log_file_write_through=ON, innodb_data_file_write_through=ON,
innodb_data_file_buffering=OFF, and
(if supported) innodb_log_file_buffering=OFF.
fsync, littlesync, nosync, or (Microsoft Windows specific) normal:
innodb_log_file_write_through=OFF, innodb_data_file_write_through=OFF,
and innodb_data_file_buffering=ON.
Note: fsync() or fdatasync() will only be disabled if the separate
parameter debug_no_sync (in the code, my_disable_sync) is set.
In mariadb-backup, the parameter innodb_flush_method will be ignored.
The Boolean parameters can be modified by SET GLOBAL while the
server is running. This will require reopening the ib_logfile0
or all currently open InnoDB data files.
We will open files straight in O_DSYNC or O_SYNC mode when applicable.
Data files we will try to open straight in O_DIRECT mode when the
page size is at least 4096 bytes. For atomically creating data files,
we will invoke os_file_set_nocache() to enable O_DIRECT afterwards,
because O_DIRECT is not supported on some file systems. We will also
continue to invoke os_file_set_nocache() on ib_logfile0 when
innodb_log_file_buffering=OFF can be fulfilled.
For reopening the ib_logfile0, we use the same logic that was developed
for online log resizing and reused for updates of
innodb_log_file_buffering.
Reopening all data files is implemented in the new function
fil_space_t::reopen_all().
Reviewed by: Vladislav Vaintroub
Tested by: Matthias Leich
2022-12-14 13:43:32 +01:00
|
|
|
{"innodb_data_file_buffering", OPT_INNODB_DATA_FILE_BUFFERING,
|
|
|
|
"Whether the file system cache for data files is enabled during --backup",
|
|
|
|
(G_PTR*) &fil_system.buffered,
|
|
|
|
(G_PTR*) &fil_system.buffered, 0, GET_BOOL, NO_ARG,
|
|
|
|
FALSE, 0, 0, 0, 0, 0},
|
|
|
|
{"innodb_data_file_write_through", OPT_INNODB_DATA_FILE_WRITE_THROUGH,
|
|
|
|
"Whether each write to data files writes through",
|
|
|
|
(G_PTR*) &fil_system.write_through,
|
|
|
|
(G_PTR*) &fil_system.write_through, 0, GET_BOOL, NO_ARG,
|
|
|
|
FALSE, 0, 0, 0, 0, 0},
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_log_file_size", OPT_INNODB_LOG_FILE_SIZE,
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
"Ignored for mysqld option compatibility",
|
|
|
|
(G_PTR*) &srv_log_file_size, (G_PTR*) &srv_log_file_size, 0,
|
2022-02-10 15:37:12 +01:00
|
|
|
GET_ULL, REQUIRED_ARG, 96 << 20, 4 << 20,
|
|
|
|
std::numeric_limits<ulonglong>::max(), 0, 4096, 0},
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_log_group_home_dir", OPT_INNODB_LOG_GROUP_HOME_DIR,
|
|
|
|
"Path to InnoDB log files.", &srv_log_group_home_dir,
|
2017-07-06 17:31:28 +02:00
|
|
|
&srv_log_group_home_dir, 0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_max_dirty_pages_pct", OPT_INNODB_MAX_DIRTY_PAGES_PCT,
|
|
|
|
"Percentage of dirty pages allowed in bufferpool.", (G_PTR*) &srv_max_buf_pool_modified_pct,
|
|
|
|
(G_PTR*) &srv_max_buf_pool_modified_pct, 0, GET_ULONG, REQUIRED_ARG, 90, 0, 100, 0, 0, 0},
|
|
|
|
{"innodb_use_native_aio", OPT_INNODB_USE_NATIVE_AIO,
|
|
|
|
"Use native AIO if supported on this platform.",
|
|
|
|
(G_PTR*) &srv_use_native_aio,
|
|
|
|
(G_PTR*) &srv_use_native_aio, 0, GET_BOOL, NO_ARG,
|
2021-11-02 14:24:20 +01:00
|
|
|
#ifdef HAVE_URING
|
|
|
|
innodb_use_native_aio_default(),
|
|
|
|
#else
|
|
|
|
TRUE,
|
|
|
|
#endif
|
|
|
|
0, 0, 0, 0, 0},
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_page_size", OPT_INNODB_PAGE_SIZE,
|
|
|
|
"The universal page size of the database.",
|
|
|
|
(G_PTR*) &innobase_page_size, (G_PTR*) &innobase_page_size, 0,
|
|
|
|
/* Use GET_LL to support numeric suffixes in 5.6 */
|
|
|
|
GET_LL, REQUIRED_ARG,
|
|
|
|
(1LL << 14), (1LL << 12), (1LL << UNIV_PAGE_SIZE_SHIFT_MAX), 0, 1L, 0},
|
|
|
|
{"innodb_buffer_pool_filename", OPT_INNODB_BUFFER_POOL_FILENAME,
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
"Ignored for mysqld option compatibility",
|
2017-04-18 20:43:20 +02:00
|
|
|
(G_PTR*) &innobase_buffer_pool_filename,
|
|
|
|
(G_PTR*) &innobase_buffer_pool_filename,
|
|
|
|
0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#ifndef DBUG_OFF /* unfortunately "debug" collides with existing options */
|
|
|
|
{"dbug", '#', "Built in DBUG debugger.",
|
|
|
|
&dbug_option, &dbug_option, 0, GET_STR, OPT_ARG,
|
|
|
|
0, 0, 0, 0, 0, 0},
|
|
|
|
#endif
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
{"innodb_checksum_algorithm", OPT_INNODB_CHECKSUM_ALGORITHM,
|
|
|
|
"The algorithm InnoDB uses for page checksumming. [CRC32, STRICT_CRC32, "
|
MDEV-25105 Remove innodb_checksum_algorithm values none,innodb,...
Historically, InnoDB supported a buggy page checksum algorithm that did not
compute a checksum over the full page. Later, well before MySQL 4.1
introduced .ibd files and the innodb_file_per_table option, the algorithm
was corrected and the first 4 bytes of each page were redefined to be
a checksum.
The original checksum was so slow that an option to disable page checksum
was introduced for benchmarketing purposes.
The Intel Nehalem microarchitecture introduced the SSE4.2 instruction set
extension, which includes instructions for faster computation of CRC-32C.
In MySQL 5.6 (and MariaDB 10.0), innodb_checksum_algorithm=crc32 was
implemented to make of that. As that option was changed to be the default
in MySQL 5.7, a bug was found on big-endian platforms and some work-around
code was added to weaken that checksum further. MariaDB disables that
work-around by default since MDEV-17958.
Later, SIMD-accelerated CRC-32C has been implemented in MariaDB for POWER
and ARM and also for IA-32/AMD64, making use of carry-less multiplication
where available.
Long story short, innodb_checksum_algorithm=crc32 is faster and more secure
than the pre-MySQL 5.6 checksum, called innodb_checksum_algorithm=innodb.
It should have removed any need to use innodb_checksum_algorithm=none.
The setting innodb_checksum_algorithm=crc32 is the default in
MySQL 5.7 and MariaDB Server 10.2, 10.3, 10.4. In MariaDB 10.5,
MDEV-19534 made innodb_checksum_algorithm=full_crc32 the default.
It is even faster and more secure.
The default settings in MariaDB do allow old data files to be read,
no matter if a worse checksum algorithm had been used.
(Unfortunately, before innodb_checksum_algorithm=full_crc32,
the data files did not identify which checksum algorithm is being used.)
The non-default settings innodb_checksum_algorithm=strict_crc32 or
innodb_checksum_algorithm=strict_full_crc32 would only allow CRC-32C
checksums. The incompatibility with old data files is why they are
not the default.
The newest server not to support innodb_checksum_algorithm=crc32
were MySQL 5.5 and MariaDB 5.5. Both have reached their end of life.
A valid reason for using innodb_checksum_algorithm=innodb could have
been the ability to downgrade. If it is really needed, data files
can be converted with an older version of the innochecksum utility.
Because there is no good reason to allow data files to be written
with insecure checksums, we will reject those option values:
innodb_checksum_algorithm=none
innodb_checksum_algorithm=innodb
innodb_checksum_algorithm=strict_none
innodb_checksum_algorithm=strict_innodb
Furthermore, the following innochecksum options will be removed,
because only strict crc32 will be supported:
innochecksum --strict-check=crc32
innochecksum -C crc32
innochecksum --write=crc32
innochecksum -w crc32
If a user wishes to convert a data file to use a different checksum
(so that it might be used with the no-longer-supported
MySQL 5.5 or MariaDB 5.5, which do not support IMPORT TABLESPACE
nor system tablespace format changes that were made in MariaDB 10.3),
then the innochecksum tool from MariaDB 10.2, 10.3, 10.4, 10.5 or
MySQL 5.7 can be used.
Reviewed by: Thirunarayanan Balathandayuthapani
2021-03-11 10:56:35 +01:00
|
|
|
"FULL_CRC32, STRICT_FULL_CRC32]", &srv_checksum_algorithm,
|
2017-04-18 20:43:20 +02:00
|
|
|
&srv_checksum_algorithm, &innodb_checksum_algorithm_typelib, GET_ENUM,
|
2018-06-14 13:23:20 +02:00
|
|
|
REQUIRED_ARG, SRV_CHECKSUM_ALGORITHM_CRC32, 0, 0, 0, 0, 0},
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
{"innodb_undo_directory", OPT_INNODB_UNDO_DIRECTORY,
|
|
|
|
"Directory where undo tablespace files live, this path can be absolute.",
|
2017-07-06 17:31:28 +02:00
|
|
|
&srv_undo_dir, &srv_undo_dir, 0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0,
|
2017-04-18 20:43:20 +02:00
|
|
|
0},
|
|
|
|
|
|
|
|
{"innodb_undo_tablespaces", OPT_INNODB_UNDO_TABLESPACES,
|
|
|
|
"Number of undo tablespaces to use.",
|
|
|
|
(G_PTR*)&srv_undo_tablespaces, (G_PTR*)&srv_undo_tablespaces,
|
2021-07-22 10:22:47 +02:00
|
|
|
0, GET_UINT, REQUIRED_ARG, 0, 0, 126, 0, 1, 0},
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2019-03-08 11:30:08 +01:00
|
|
|
{"innodb_compression_level", OPT_INNODB_COMPRESSION_LEVEL,
|
|
|
|
"Compression level used for zlib compression.",
|
|
|
|
(G_PTR*)&page_zip_level, (G_PTR*)&page_zip_level,
|
|
|
|
0, GET_UINT, REQUIRED_ARG, 6, 0, 9, 0, 0, 0},
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
{"defaults_group", OPT_DEFAULTS_GROUP, "defaults group in config file (default \"mysqld\").",
|
|
|
|
(G_PTR*) &defaults_group, (G_PTR*) &defaults_group,
|
|
|
|
0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
2017-12-18 16:37:06 +01:00
|
|
|
{"plugin-dir", OPT_PLUGIN_DIR,
|
2021-10-08 09:48:31 +02:00
|
|
|
"Server plugin directory. Used to load plugins during 'prepare' phase."
|
2017-12-18 16:37:06 +01:00
|
|
|
"Has no effect in the 'backup' phase (plugin directory during backup is the same as server's)",
|
2017-04-18 21:05:57 +02:00
|
|
|
&xb_plugin_dir, &xb_plugin_dir,
|
|
|
|
0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0 },
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
{"open_files_limit", OPT_OPEN_FILES_LIMIT, "the maximum number of file "
|
|
|
|
"descriptors to reserve with setrlimit().",
|
|
|
|
(G_PTR*) &xb_open_files_limit, (G_PTR*) &xb_open_files_limit, 0, GET_ULONG,
|
|
|
|
REQUIRED_ARG, 0, 0, UINT_MAX, 0, 1, 0},
|
|
|
|
|
2017-09-11 18:45:36 +02:00
|
|
|
{"lock-ddl-per-table", OPT_LOCK_DDL_PER_TABLE, "Lock DDL for each table "
|
2021-05-31 12:27:47 +02:00
|
|
|
"before backup starts to copy it and until the backup is completed.",
|
2017-09-11 18:45:36 +02:00
|
|
|
(uchar*) &opt_lock_ddl_per_table, (uchar*) &opt_lock_ddl_per_table, 0,
|
|
|
|
GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
2018-06-07 16:13:54 +02:00
|
|
|
{"rocksdb-datadir", OPT_ROCKSDB_DATADIR, "RocksDB data directory."
|
|
|
|
"This option is only used with --copy-back or --move-back option",
|
|
|
|
&xb_rocksdb_datadir, &xb_rocksdb_datadir,
|
|
|
|
0, GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0 },
|
|
|
|
|
2022-01-18 19:07:01 +01:00
|
|
|
{"rocksdb-backup", OPT_BACKUP_ROCKSDB, "Backup rocksdb data, if rocksdb plugin is installed."
|
2018-06-07 16:13:54 +02:00
|
|
|
"Used only with --backup option. Can be useful for partial backups, to exclude all rocksdb data",
|
|
|
|
&xb_backup_rocksdb, &xb_backup_rocksdb,
|
|
|
|
0, GET_BOOL, NO_ARG, 1, 0, 0, 0, 0, 0 },
|
|
|
|
|
2022-01-18 19:07:01 +01:00
|
|
|
{"check-privileges", OPT_XTRA_CHECK_PRIVILEGES, "Check database user "
|
2018-12-14 23:36:21 +01:00
|
|
|
"privileges fro the backup user",
|
|
|
|
&opt_check_privileges, &opt_check_privileges,
|
|
|
|
0, GET_BOOL, NO_ARG, 1, 0, 0, 0, 0, 0 },
|
2018-06-07 16:13:54 +02:00
|
|
|
|
2021-04-01 11:33:59 +02:00
|
|
|
{"innodb_force_recovery", OPT_INNODB_FORCE_RECOVERY,
|
|
|
|
"(for --prepare): Crash recovery mode (ignores "
|
|
|
|
"page corruption; for emergencies only).",
|
|
|
|
(G_PTR*)&srv_force_recovery,
|
|
|
|
(G_PTR*)&srv_force_recovery,
|
|
|
|
0, GET_ULONG, OPT_ARG, 0, 0, SRV_FORCE_IGNORE_CORRUPT, 0, 0, 0},
|
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
{"mysqld-args", OPT_XTRA_MYSQLD_ARGS,
|
|
|
|
"All arguments that follow this argument are considered as server "
|
|
|
|
"options, and if some of them are not supported by mariabackup, they "
|
|
|
|
"will be ignored.",
|
|
|
|
(G_PTR *) &xtrabackup_mysqld_args, (G_PTR *) &xtrabackup_mysqld_args, 0,
|
|
|
|
GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
2020-06-23 17:07:03 +02:00
|
|
|
{"help", '?',
|
|
|
|
"Display this help and exit.",
|
2020-06-23 17:07:03 +02:00
|
|
|
(G_PTR *) &xtrabackup_help, (G_PTR *) &xtrabackup_help, 0,
|
2020-06-23 17:07:03 +02:00
|
|
|
GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0},
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
{ 0, 0, 0, 0, 0, 0, GET_NO_ARG, NO_ARG, 0, 0, 0, 0, 0, 0}
|
|
|
|
};
|
|
|
|
|
|
|
|
uint xb_server_options_count = array_elements(xb_server_options);
|
|
|
|
|
2017-09-08 00:53:21 +02:00
|
|
|
|
2018-12-21 11:18:53 +01:00
|
|
|
static std::set<std::string> tables_for_export;
|
2017-09-08 00:53:21 +02:00
|
|
|
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
static void append_export_table(const char *dbname, const char *tablename,
|
2021-05-17 17:12:33 +02:00
|
|
|
bool is_remote, bool skip_node_page0,
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t defer_space_id)
|
2017-09-08 00:53:21 +02:00
|
|
|
{
|
|
|
|
if(dbname && tablename && !is_remote)
|
|
|
|
{
|
|
|
|
char buf[3*FN_REFLEN];
|
|
|
|
snprintf(buf,sizeof(buf),"%s/%s",dbname, tablename);
|
|
|
|
// trim .ibd
|
|
|
|
char *p=strrchr(buf, '.');
|
|
|
|
if (p) *p=0;
|
|
|
|
|
2018-12-21 11:18:53 +01:00
|
|
|
std::string name=ut_get_name(0, buf);
|
|
|
|
/* Strip partition name comment from table name, if any */
|
|
|
|
if (ends_with(name.c_str(), "*/"))
|
|
|
|
{
|
|
|
|
size_t pos= name.rfind("/*");
|
|
|
|
if (pos != std::string::npos)
|
|
|
|
name.resize(pos);
|
|
|
|
}
|
|
|
|
tables_for_export.insert(name);
|
2017-09-08 00:53:21 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
#define BOOTSTRAP_FILENAME "mariabackup_prepare_for_export.sql"
|
|
|
|
|
|
|
|
static int create_bootstrap_file()
|
|
|
|
{
|
|
|
|
FILE *f= fopen(BOOTSTRAP_FILENAME,"wb");
|
|
|
|
if(!f)
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
fputs("SET NAMES UTF8;\n",f);
|
|
|
|
enumerate_ibd_files(append_export_table);
|
2018-12-21 11:18:53 +01:00
|
|
|
for (std::set<std::string>::iterator it = tables_for_export.begin();
|
|
|
|
it != tables_for_export.end(); it++)
|
2017-09-08 00:53:21 +02:00
|
|
|
{
|
2018-12-21 11:18:53 +01:00
|
|
|
const char *tab = it->c_str();
|
2017-09-08 00:53:21 +02:00
|
|
|
fprintf(f,
|
|
|
|
"BEGIN NOT ATOMIC "
|
|
|
|
"DECLARE CONTINUE HANDLER FOR NOT FOUND,SQLEXCEPTION BEGIN END;"
|
|
|
|
"FLUSH TABLES %s FOR EXPORT;"
|
|
|
|
"END;\n"
|
|
|
|
"UNLOCK TABLES;\n",
|
|
|
|
tab);
|
|
|
|
}
|
|
|
|
fclose(f);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int prepare_export()
|
|
|
|
{
|
|
|
|
int err= -1;
|
|
|
|
|
|
|
|
char cmdline[2*FN_REFLEN];
|
|
|
|
FILE *outf;
|
|
|
|
|
|
|
|
if (create_bootstrap_file())
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
// Process defaults-file , it can have some --lc-language stuff,
|
|
|
|
// which is* unfortunately* still necessary to get mysqld up
|
2021-04-07 16:53:23 +02:00
|
|
|
if (strncmp(orig_argv1,"--defaults-file=", 16) == 0)
|
2017-09-08 00:53:21 +02:00
|
|
|
{
|
2018-08-03 10:24:05 +02:00
|
|
|
snprintf(cmdline, sizeof cmdline,
|
2021-04-07 16:53:23 +02:00
|
|
|
IF_WIN("\"","") "\"%s\" --mysqld \"%s\""
|
2018-01-16 04:24:27 +01:00
|
|
|
" --defaults-extra-file=./backup-my.cnf --defaults-group-suffix=%s --datadir=."
|
2018-12-21 11:18:53 +01:00
|
|
|
" --innodb --innodb-fast-shutdown=0 --loose-partition"
|
2017-09-08 00:53:21 +02:00
|
|
|
" --innodb_purge_rseg_truncate_frequency=1 --innodb-buffer-pool-size=%llu"
|
2021-04-07 16:53:23 +02:00
|
|
|
" --console --skip-log-error --skip-log-bin --bootstrap %s< "
|
2019-10-01 02:23:33 +02:00
|
|
|
BOOTSTRAP_FILENAME IF_WIN("\"",""),
|
2021-04-01 11:33:59 +02:00
|
|
|
mariabackup_exe,
|
2017-09-08 00:53:21 +02:00
|
|
|
orig_argv1, (my_defaults_group_suffix?my_defaults_group_suffix:""),
|
2021-04-01 11:33:59 +02:00
|
|
|
xtrabackup_use_memory,
|
2021-04-07 16:53:23 +02:00
|
|
|
(srv_force_recovery ? "--innodb-force-recovery=1 " : ""));
|
2017-09-08 00:53:21 +02:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2021-04-01 11:33:59 +02:00
|
|
|
snprintf(cmdline, sizeof cmdline,
|
2021-04-07 16:53:23 +02:00
|
|
|
IF_WIN("\"","") "\"%s\" --mysqld"
|
2018-01-16 04:24:27 +01:00
|
|
|
" --defaults-file=./backup-my.cnf --defaults-group-suffix=%s --datadir=."
|
2018-12-21 11:18:53 +01:00
|
|
|
" --innodb --innodb-fast-shutdown=0 --loose-partition"
|
2017-09-08 00:53:21 +02:00
|
|
|
" --innodb_purge_rseg_truncate_frequency=1 --innodb-buffer-pool-size=%llu"
|
2021-04-07 16:53:23 +02:00
|
|
|
" --console --log-error= --skip-log-bin --bootstrap %s< "
|
2019-10-01 02:23:33 +02:00
|
|
|
BOOTSTRAP_FILENAME IF_WIN("\"",""),
|
2017-09-08 00:53:21 +02:00
|
|
|
mariabackup_exe,
|
2018-01-16 04:24:27 +01:00
|
|
|
(my_defaults_group_suffix?my_defaults_group_suffix:""),
|
2021-04-01 11:33:59 +02:00
|
|
|
xtrabackup_use_memory,
|
2021-04-07 16:53:23 +02:00
|
|
|
(srv_force_recovery ? "--innodb-force-recovery=1 " : ""));
|
2017-09-08 00:53:21 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
msg("Prepare export : executing %s\n", cmdline);
|
|
|
|
fflush(stderr);
|
|
|
|
|
|
|
|
outf= popen(cmdline,"r");
|
|
|
|
if (!outf)
|
|
|
|
goto end;
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
|
2017-09-08 00:53:21 +02:00
|
|
|
char outline[FN_REFLEN];
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
while (fgets(outline, FN_REFLEN - 1, outf))
|
2017-09-08 00:53:21 +02:00
|
|
|
fprintf(stderr,"%s",outline);
|
|
|
|
|
|
|
|
err = pclose(outf);
|
|
|
|
end:
|
|
|
|
unlink(BOOTSTRAP_FILENAME);
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
static const char *xb_client_default_groups[]= {
|
2020-06-16 11:02:13 +02:00
|
|
|
"client", "client-server", "client-mariadb", 0, 0, 0};
|
2017-09-08 00:53:21 +02:00
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
static const char *backup_default_groups[]= {
|
|
|
|
"xtrabackup", "mariabackup", "mariadb-backup", 0, 0, 0};
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
static void print_version(void)
|
|
|
|
{
|
2019-02-21 00:06:08 +01:00
|
|
|
fprintf(stderr, "%s based on MariaDB server %s %s (%s)\n",
|
2017-04-18 21:05:57 +02:00
|
|
|
my_progname, MYSQL_SERVER_VERSION, SYSTEM_TYPE, MACHINE_TYPE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
static void usage(void)
|
|
|
|
{
|
|
|
|
puts("Open source backup tool for InnoDB and XtraDB\n\
|
|
|
|
\n\
|
|
|
|
Copyright (C) 2009-2015 Percona LLC and/or its affiliates.\n\
|
|
|
|
Portions Copyright (C) 2000, 2011, MySQL AB & Innobase Oy. All Rights Reserved.\n\
|
|
|
|
\n\
|
|
|
|
This program is free software; you can redistribute it and/or\n\
|
|
|
|
modify it under the terms of the GNU General Public License\n\
|
|
|
|
as published by the Free Software Foundation version 2\n\
|
|
|
|
of the License.\n\
|
|
|
|
\n\
|
|
|
|
This program is distributed in the hope that it will be useful,\n\
|
|
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of\n\
|
|
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the\n\
|
|
|
|
GNU General Public License for more details.\n\
|
|
|
|
\n\
|
|
|
|
You can download full text of the license on http://www.gnu.org/licenses/gpl-2.0.txt\n");
|
|
|
|
|
2018-01-22 16:10:52 +01:00
|
|
|
printf("Usage: %s [--defaults-file=#] [--backup | --prepare | --copy-back | --move-back] [OPTIONS]\n",my_progname);
|
2020-06-08 10:45:56 +02:00
|
|
|
print_defaults("my", load_default_groups);
|
2017-04-18 20:43:20 +02:00
|
|
|
my_print_help(xb_client_options);
|
|
|
|
my_print_help(xb_server_options);
|
|
|
|
my_print_variables(xb_server_options);
|
|
|
|
my_print_variables(xb_client_options);
|
|
|
|
}
|
|
|
|
|
|
|
|
#define ADD_PRINT_PARAM_OPT(value) \
|
|
|
|
{ \
|
|
|
|
print_param_str << opt->name << "=" << value << "\n"; \
|
|
|
|
param_set.insert(opt->name); \
|
|
|
|
}
|
|
|
|
|
|
|
|
/************************************************************************
|
|
|
|
Check if parameter is set in defaults file or via command line argument
|
|
|
|
@return true if parameter is set. */
|
|
|
|
bool
|
|
|
|
check_if_param_set(const char *param)
|
|
|
|
{
|
|
|
|
return param_set.find(param) != param_set.end();
|
|
|
|
}
|
|
|
|
|
|
|
|
my_bool
|
2019-09-29 16:30:57 +02:00
|
|
|
xb_get_one_option(const struct my_option *opt,
|
2021-02-05 13:57:46 +01:00
|
|
|
const char *argument, const char *)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2019-09-29 16:30:57 +02:00
|
|
|
switch(opt->id) {
|
2017-04-18 20:43:20 +02:00
|
|
|
case 'h':
|
|
|
|
strmake(mysql_real_data_home,argument, FN_REFLEN - 1);
|
|
|
|
mysql_data_home= mysql_real_data_home;
|
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(mysql_real_data_home);
|
|
|
|
break;
|
|
|
|
|
|
|
|
case 't':
|
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(opt_mysql_tmpdir);
|
|
|
|
break;
|
|
|
|
|
|
|
|
case OPT_INNODB_DATA_HOME_DIR:
|
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(innobase_data_home_dir);
|
|
|
|
break;
|
|
|
|
|
|
|
|
case OPT_INNODB_DATA_FILE_PATH:
|
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(innobase_data_file_path);
|
|
|
|
break;
|
|
|
|
|
|
|
|
case OPT_INNODB_LOG_GROUP_HOME_DIR:
|
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(srv_log_group_home_dir);
|
|
|
|
break;
|
|
|
|
|
|
|
|
case OPT_INNODB_PAGE_SIZE:
|
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(innobase_page_size);
|
|
|
|
break;
|
|
|
|
|
|
|
|
case OPT_INNODB_UNDO_DIRECTORY:
|
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(srv_undo_dir);
|
|
|
|
break;
|
|
|
|
|
|
|
|
case OPT_INNODB_UNDO_TABLESPACES:
|
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(srv_undo_tablespaces);
|
|
|
|
break;
|
|
|
|
|
|
|
|
case OPT_INNODB_CHECKSUM_ALGORITHM:
|
|
|
|
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
ut_a(srv_checksum_algorithm <= SRV_CHECKSUM_ALGORITHM_STRICT_FULL_CRC32);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(innodb_checksum_algorithm_names[srv_checksum_algorithm]);
|
|
|
|
break;
|
|
|
|
|
2019-03-08 11:30:08 +01:00
|
|
|
case OPT_INNODB_COMPRESSION_LEVEL:
|
|
|
|
ADD_PRINT_PARAM_OPT(page_zip_level);
|
|
|
|
break;
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
case OPT_INNODB_BUFFER_POOL_FILENAME:
|
|
|
|
|
|
|
|
ADD_PRINT_PARAM_OPT(innobase_buffer_pool_filename);
|
|
|
|
break;
|
|
|
|
|
2021-04-01 11:33:59 +02:00
|
|
|
case OPT_INNODB_FORCE_RECOVERY:
|
|
|
|
|
|
|
|
if (srv_force_recovery) {
|
|
|
|
ADD_PRINT_PARAM_OPT(srv_force_recovery);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
case OPT_XTRA_TARGET_DIR:
|
|
|
|
strmake(xtrabackup_real_target_dir,argument, sizeof(xtrabackup_real_target_dir)-1);
|
|
|
|
xtrabackup_target_dir= xtrabackup_real_target_dir;
|
|
|
|
break;
|
|
|
|
case OPT_XTRA_STREAM:
|
2020-01-14 13:57:14 +01:00
|
|
|
if (!strcasecmp(argument, "mbstream") ||
|
|
|
|
!strcasecmp(argument, "xbstream"))
|
2017-04-18 20:43:20 +02:00
|
|
|
xtrabackup_stream_fmt = XB_STREAM_FMT_XBSTREAM;
|
|
|
|
else
|
|
|
|
{
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Invalid --stream argument: %s", argument);
|
2017-04-18 20:43:20 +02:00
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
xtrabackup_stream = TRUE;
|
|
|
|
break;
|
|
|
|
case OPT_XTRA_COMPRESS:
|
|
|
|
if (argument == NULL)
|
|
|
|
xtrabackup_compress_alg = "quicklz";
|
|
|
|
else if (strcasecmp(argument, "quicklz"))
|
|
|
|
{
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Invalid --compress argument: %s", argument);
|
2017-04-18 20:43:20 +02:00
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
xtrabackup_compress = TRUE;
|
|
|
|
break;
|
|
|
|
case OPT_DECOMPRESS:
|
|
|
|
opt_decompress = TRUE;
|
|
|
|
xtrabackup_decrypt_decompress = true;
|
|
|
|
break;
|
|
|
|
case (int) OPT_CORE_FILE:
|
|
|
|
test_flags |= TEST_CORE_ON_SIGNAL;
|
|
|
|
break;
|
|
|
|
case OPT_HISTORY:
|
|
|
|
if (argument) {
|
|
|
|
opt_history = argument;
|
|
|
|
} else {
|
|
|
|
opt_history = "";
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case 'p':
|
2020-06-08 10:45:56 +02:00
|
|
|
opt_password = argument;
|
2017-04-18 20:43:20 +02:00
|
|
|
break;
|
2017-08-21 19:16:12 +02:00
|
|
|
case OPT_PROTOCOL:
|
|
|
|
if (argument)
|
|
|
|
{
|
2017-12-18 14:15:48 +01:00
|
|
|
if ((opt_protocol= find_type_with_warning(argument, &sql_protocol_typelib,
|
|
|
|
opt->name)) <= 0)
|
|
|
|
{
|
|
|
|
sf_leaking_memory= 1; /* no memory leak reports here */
|
|
|
|
exit(1);
|
|
|
|
}
|
2017-08-21 19:16:12 +02:00
|
|
|
}
|
|
|
|
break;
|
2018-05-08 16:08:08 +02:00
|
|
|
#define MYSQL_CLIENT
|
2017-04-18 20:43:20 +02:00
|
|
|
#include "sslopt-case.h"
|
2018-05-08 16:08:08 +02:00
|
|
|
#undef MYSQL_CLIENT
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
case '?':
|
|
|
|
usage();
|
|
|
|
exit(EXIT_SUCCESS);
|
|
|
|
break;
|
|
|
|
case 'v':
|
|
|
|
print_version();
|
|
|
|
exit(EXIT_SUCCESS);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2018-04-29 08:41:42 +02:00
|
|
|
static bool innodb_init_param()
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2017-04-18 21:05:57 +02:00
|
|
|
srv_is_being_started = TRUE;
|
2017-04-18 20:43:20 +02:00
|
|
|
/* === some variables from mysqld === */
|
|
|
|
memset((G_PTR) &mysql_tmpdir_list, 0, sizeof(mysql_tmpdir_list));
|
|
|
|
|
2019-01-17 11:33:31 +01:00
|
|
|
if (init_tmpdir(&mysql_tmpdir_list, opt_mysql_tmpdir)) {
|
|
|
|
msg("init_tmpdir() failed");
|
2018-04-29 08:41:42 +02:00
|
|
|
return true;
|
2019-01-17 11:33:31 +01:00
|
|
|
}
|
2017-04-18 21:05:57 +02:00
|
|
|
xtrabackup_tmpdir = my_tmpdir(&mysql_tmpdir_list);
|
2017-04-18 20:43:20 +02:00
|
|
|
/* dummy for initialize all_charsets[] */
|
|
|
|
get_charset_name(0);
|
|
|
|
|
|
|
|
srv_page_size = 0;
|
|
|
|
srv_page_size_shift = 0;
|
2018-01-16 12:42:33 +01:00
|
|
|
#ifdef BTR_CUR_HASH_ADAPT
|
|
|
|
btr_ahi_parts = 1;
|
|
|
|
#endif /* BTR_CUR_HASH_ADAPT */
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (innobase_page_size != (1LL << 14)) {
|
2018-04-27 13:26:43 +02:00
|
|
|
size_t n_shift = get_bit_shift(size_t(innobase_page_size));
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (n_shift >= 12 && n_shift <= UNIV_PAGE_SIZE_SHIFT_MAX) {
|
2021-07-22 10:22:47 +02:00
|
|
|
srv_page_size_shift = uint32_t(n_shift);
|
2018-04-27 13:26:43 +02:00
|
|
|
srv_page_size = 1U << n_shift;
|
2017-04-18 20:43:20 +02:00
|
|
|
msg("InnoDB: The universal page size of the "
|
2019-01-14 22:28:23 +01:00
|
|
|
"database is set to %lu.", srv_page_size);
|
2017-04-18 20:43:20 +02:00
|
|
|
} else {
|
2019-01-17 11:33:31 +01:00
|
|
|
msg("invalid value of "
|
2017-04-18 20:43:20 +02:00
|
|
|
"innobase_page_size: %lld", innobase_page_size);
|
2018-04-29 08:41:42 +02:00
|
|
|
goto error;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
srv_page_size_shift = 14;
|
2018-04-27 13:26:43 +02:00
|
|
|
srv_page_size = 1U << 14;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Check that values don't overflow on 32-bit systems. */
|
|
|
|
if (sizeof(ulint) == 4) {
|
|
|
|
if (xtrabackup_use_memory > UINT_MAX32) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: use-memory can't be over 4GB"
|
2019-01-14 22:28:23 +01:00
|
|
|
" on 32-bit systems");
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
static char default_path[2] = { FN_CURLIB, 0 };
|
|
|
|
fil_path_to_mysql_datadir = default_path;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* Set InnoDB initialization parameters according to the values
|
|
|
|
read from MySQL .cnf file */
|
|
|
|
|
2017-05-29 13:24:18 +02:00
|
|
|
if (xtrabackup_backup) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: using the following InnoDB configuration:");
|
2017-04-18 20:43:20 +02:00
|
|
|
} else {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: using the following InnoDB configuration "
|
2019-01-14 22:28:23 +01:00
|
|
|
"for recovery:");
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/*--------------- Data files -------------------------*/
|
|
|
|
|
|
|
|
/* The default dir for data files is the datadir of MySQL */
|
|
|
|
|
2017-05-29 13:24:18 +02:00
|
|
|
srv_data_home = (xtrabackup_backup && innobase_data_home_dir
|
2017-04-18 20:43:20 +02:00
|
|
|
? innobase_data_home_dir : default_path);
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("innodb_data_home_dir = %s", srv_data_home);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* Set default InnoDB data file size to 10 MB and let it be
|
|
|
|
auto-extending. Thus users can use InnoDB in >= 4.0 without having
|
|
|
|
to specify any startup options. */
|
|
|
|
|
|
|
|
if (!innobase_data_file_path) {
|
|
|
|
innobase_data_file_path = (char*) "ibdata1:10M:autoextend";
|
|
|
|
}
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("innodb_data_file_path = %s",
|
2017-04-18 20:43:20 +02:00
|
|
|
innobase_data_file_path);
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
srv_sys_space.set_space_id(TRX_SYS_SPACE);
|
|
|
|
srv_sys_space.set_path(srv_data_home);
|
MDEV-12026: Implement innodb_checksum_algorithm=full_crc32
MariaDB data-at-rest encryption (innodb_encrypt_tables)
had repurposed the same unused data field that was repurposed
in MySQL 5.7 (and MariaDB 10.2) for the Split Sequence Number (SSN)
field of SPATIAL INDEX. Because of this, MariaDB was unable to
support encryption on SPATIAL INDEX pages.
Furthermore, InnoDB page checksums skipped some bytes, and there
are multiple variations and checksum algorithms. By default,
InnoDB accepts all variations of all algorithms that ever existed.
This unnecessarily weakens the page checksums.
We hereby introduce two more innodb_checksum_algorithm variants
(full_crc32, strict_full_crc32) that are special in a way:
When either setting is active, newly created data files will
carry a flag (fil_space_t::full_crc32()) that indicates that
all pages of the file will use a full CRC-32C checksum over the
entire page contents (excluding the bytes where the checksum
is stored, at the very end of the page). Such files will always
use that checksum, no matter what the parameter
innodb_checksum_algorithm is assigned to.
For old files, the old checksum algorithms will continue to be
used. The value strict_full_crc32 will be equivalent to strict_crc32
and the value full_crc32 will be equivalent to crc32.
ROW_FORMAT=COMPRESSED tables will only use the old format.
These tables do not support new features, such as larger
innodb_page_size or instant ADD/DROP COLUMN. They may be
deprecated in the future. We do not want an unnecessary
file format change for them.
The new full_crc32() format also cleans up the MariaDB tablespace
flags. We will reserve flags to store the page_compressed
compression algorithm, and to store the compressed payload length,
so that checksum can be computed over the compressed (and
possibly encrypted) stream and can be validated without
decrypting or decompressing the page.
In the full_crc32 format, there no longer are separate before-encryption
and after-encryption checksums for pages. The single checksum is
computed on the page contents that is written to the file.
We do not make the new algorithm the default for two reasons.
First, MariaDB 10.4.2 was a beta release, and the default values
of parameters should not change after beta. Second, we did not
yet implement the full_crc32 format for page_compressed pages.
This will be fixed in MDEV-18644.
This is joint work with Marko Mäkelä.
2019-02-19 20:00:00 +01:00
|
|
|
switch (srv_checksum_algorithm) {
|
|
|
|
case SRV_CHECKSUM_ALGORITHM_FULL_CRC32:
|
|
|
|
case SRV_CHECKSUM_ALGORITHM_STRICT_FULL_CRC32:
|
|
|
|
srv_sys_space.set_flags(FSP_FLAGS_FCRC32_MASK_MARKER
|
|
|
|
| FSP_FLAGS_FCRC32_PAGE_SSIZE());
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
srv_sys_space.set_flags(FSP_FLAGS_PAGE_SSIZE());
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (!srv_sys_space.parse_params(innobase_data_file_path, true)) {
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
2018-04-29 08:41:42 +02:00
|
|
|
srv_sys_space.normalize_size();
|
|
|
|
srv_lock_table_size = 5 * (srv_buf_pool_size >> srv_page_size_shift);
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* -------------- Log files ---------------------------*/
|
|
|
|
|
|
|
|
/* The default dir for log files is the datadir of MySQL */
|
|
|
|
|
2017-05-29 13:24:18 +02:00
|
|
|
if (!(xtrabackup_backup && srv_log_group_home_dir)) {
|
2017-04-18 20:43:20 +02:00
|
|
|
srv_log_group_home_dir = default_path;
|
|
|
|
}
|
|
|
|
if (xtrabackup_prepare && xtrabackup_incremental_dir) {
|
|
|
|
srv_log_group_home_dir = xtrabackup_incremental_dir;
|
|
|
|
}
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("innodb_log_group_home_dir = %s",
|
2017-04-18 20:43:20 +02:00
|
|
|
srv_log_group_home_dir);
|
|
|
|
|
|
|
|
if (strchr(srv_log_group_home_dir, ';')) {
|
|
|
|
msg("syntax error in innodb_log_group_home_dir, ");
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
goto error;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
srv_adaptive_flushing = FALSE;
|
|
|
|
|
|
|
|
/* We set srv_pool_size here in units of 1 kB. InnoDB internally
|
|
|
|
changes the value so that it becomes the number of database pages. */
|
|
|
|
|
|
|
|
srv_buf_pool_size = (ulint) xtrabackup_use_memory;
|
2021-12-02 04:30:26 +01:00
|
|
|
srv_buf_pool_chunk_unit = srv_buf_pool_size;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-12-14 14:27:03 +01:00
|
|
|
srv_n_read_io_threads = (uint) innobase_read_io_threads;
|
|
|
|
srv_n_write_io_threads = (uint) innobase_write_io_threads;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2018-04-29 08:41:42 +02:00
|
|
|
srv_max_n_open_files = ULINT_UNDEFINED - 5;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2018-06-15 13:09:15 +02:00
|
|
|
srv_print_verbose_log = verbose ? 2 : 1;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2018-04-29 08:41:42 +02:00
|
|
|
ut_ad(DATA_MYSQL_BINARY_CHARSET_COLL == my_charset_bin.number);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#ifdef _WIN32
|
|
|
|
srv_use_native_aio = TRUE;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
#elif defined(LINUX_NATIVE_AIO)
|
|
|
|
|
|
|
|
if (srv_use_native_aio) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("InnoDB: Using Linux native AIO");
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2021-03-15 10:30:17 +01:00
|
|
|
#elif defined(HAVE_URING)
|
2021-11-02 14:24:20 +01:00
|
|
|
if (!srv_use_native_aio) {
|
|
|
|
} else if (io_uring_may_be_unsafe) {
|
|
|
|
msg("InnoDB: Using liburing on this kernel %s may cause hangs;"
|
|
|
|
" see https://jira.mariadb.org/browse/MDEV-26674",
|
|
|
|
io_uring_may_be_unsafe);
|
|
|
|
} else {
|
2021-03-15 10:30:17 +01:00
|
|
|
msg("InnoDB: Using liburing");
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
#else
|
|
|
|
/* Currently native AIO is supported only on windows and linux
|
|
|
|
and that also when the support is compiled in. In all other
|
|
|
|
cases, we ignore the setting of innodb_use_native_aio. */
|
|
|
|
srv_use_native_aio = FALSE;
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* Assign the default value to srv_undo_dir if it's not specified, as
|
|
|
|
my_getopt does not support default values for string options. We also
|
|
|
|
ignore the option and override innodb_undo_directory on --prepare,
|
|
|
|
because separate undo tablespaces are copied to the root backup
|
|
|
|
directory. */
|
|
|
|
|
|
|
|
if (!srv_undo_dir || !xtrabackup_backup) {
|
2017-07-06 17:31:28 +02:00
|
|
|
srv_undo_dir = (char*) ".";
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2021-04-01 11:33:59 +02:00
|
|
|
compile_time_assert(SRV_FORCE_IGNORE_CORRUPT == 1);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This option can be read both from the command line, and the
|
|
|
|
* defaults file. The assignment should account for both cases,
|
|
|
|
* and for "--innobackupex". Since the command line argument is
|
|
|
|
* parsed after the defaults file, it takes precedence.
|
|
|
|
*/
|
|
|
|
if (xtrabackup_innodb_force_recovery) {
|
|
|
|
srv_force_recovery = xtrabackup_innodb_force_recovery;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (srv_force_recovery >= SRV_FORCE_IGNORE_CORRUPT) {
|
|
|
|
if (!xtrabackup_prepare) {
|
|
|
|
msg("mariabackup: The option \"innodb_force_recovery\""
|
|
|
|
" should only be used with \"%s\".",
|
|
|
|
(innobackupex_mode ? "--apply-log" : "--prepare"));
|
|
|
|
goto error;
|
|
|
|
} else {
|
|
|
|
msg("innodb_force_recovery = %lu", srv_force_recovery);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-04-29 08:41:42 +02:00
|
|
|
#ifdef _WIN32
|
|
|
|
srv_use_native_aio = TRUE;
|
|
|
|
#endif
|
|
|
|
return false;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
error:
|
2018-12-21 21:06:08 +01:00
|
|
|
msg("mariabackup: innodb_init_param(): Error occurred.\n");
|
2018-04-29 08:41:42 +02:00
|
|
|
return true;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
static byte log_hdr_buf[log_t::START_OFFSET + SIZE_OF_FILE_CHECKPOINT];
|
|
|
|
|
|
|
|
/** Initialize an InnoDB log file header in log_hdr_buf[] */
|
|
|
|
static void log_hdr_init()
|
|
|
|
{
|
|
|
|
memset(log_hdr_buf, 0, sizeof log_hdr_buf);
|
|
|
|
mach_write_to_4(LOG_HEADER_FORMAT + log_hdr_buf, log_t::FORMAT_10_8);
|
|
|
|
mach_write_to_8(LOG_HEADER_START_LSN + log_hdr_buf,
|
|
|
|
log_sys.next_checkpoint_lsn);
|
|
|
|
snprintf(reinterpret_cast<char*>(LOG_HEADER_CREATOR + log_hdr_buf),
|
|
|
|
16, "Backup %u.%u.%u",
|
|
|
|
MYSQL_VERSION_ID / 10000, MYSQL_VERSION_ID / 100 % 100,
|
|
|
|
MYSQL_VERSION_ID % 100);
|
|
|
|
if (log_sys.is_encrypted())
|
|
|
|
log_crypt_write_header(log_hdr_buf + LOG_HEADER_CREATOR_END);
|
|
|
|
mach_write_to_4(508 + log_hdr_buf, my_crc32c(0, log_hdr_buf, 508));
|
|
|
|
mach_write_to_8(log_hdr_buf + 0x1000, log_sys.next_checkpoint_lsn);
|
|
|
|
mach_write_to_8(log_hdr_buf + 0x1008, recv_sys.lsn);
|
|
|
|
mach_write_to_4(log_hdr_buf + 0x103c,
|
|
|
|
my_crc32c(0, log_hdr_buf + 0x1000, 60));
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
static bool innodb_init()
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
bool create_new_db= false;
|
2021-01-06 12:53:14 +01:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
srv_max_io_capacity= srv_io_capacity >= SRV_MAX_IO_CAPACITY_LIMIT / 2
|
|
|
|
? SRV_MAX_IO_CAPACITY_LIMIT : std::max(2 * srv_io_capacity, 2000UL);
|
2021-01-06 12:53:14 +01:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
/* Check if the data files exist or not. */
|
|
|
|
dberr_t err= srv_sys_space.check_file_spec(&create_new_db, 5U << 20);
|
2018-04-29 08:41:42 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (err == DB_SUCCESS)
|
|
|
|
err= srv_start(create_new_db);
|
2018-04-29 08:41:42 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (err != DB_SUCCESS)
|
|
|
|
{
|
|
|
|
msg("mariadb-backup: srv_start() returned %d (%s).", err, ut_strerr(err));
|
|
|
|
return true;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
ut_ad(srv_force_recovery <= SRV_FORCE_IGNORE_CORRUPT);
|
|
|
|
ut_ad(recv_no_log_write);
|
|
|
|
buf_flush_sync();
|
|
|
|
DBUG_ASSERT(!buf_pool.any_io_pending());
|
|
|
|
log_sys.close_file();
|
|
|
|
|
|
|
|
if (xtrabackup_incremental)
|
|
|
|
/* Reset the ib_logfile0 in --target-dir, not --incremental-dir. */
|
|
|
|
srv_log_group_home_dir= xtrabackup_target_dir;
|
|
|
|
|
|
|
|
bool ret;
|
|
|
|
const std::string ib_logfile0{get_log_file_path()};
|
2022-02-15 14:03:15 +01:00
|
|
|
os_file_delete_if_exists_func(ib_logfile0.c_str(), nullptr);
|
|
|
|
os_file_t file= os_file_create_func(ib_logfile0.c_str(),
|
|
|
|
OS_FILE_CREATE, OS_FILE_NORMAL,
|
|
|
|
OS_DATA_FILE_NO_O_DIRECT, false, &ret);
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (!ret)
|
|
|
|
{
|
|
|
|
invalid_log:
|
|
|
|
msg("mariadb-backup: Cannot create %s", ib_logfile0.c_str());
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
recv_sys.lsn= log_sys.next_checkpoint_lsn=
|
|
|
|
log_sys.get_lsn() - SIZE_OF_FILE_CHECKPOINT;
|
|
|
|
log_sys.set_latest_format(false); // not encrypted
|
|
|
|
log_hdr_init();
|
|
|
|
byte *b= &log_hdr_buf[log_t::START_OFFSET];
|
|
|
|
b[0]= FILE_CHECKPOINT | 10;
|
|
|
|
mach_write_to_8(b + 3, recv_sys.lsn);
|
|
|
|
b[11]= 1;
|
|
|
|
mach_write_to_4(b + 12, my_crc32c(0, b, 11));
|
|
|
|
static_assert(12 + 4 == SIZE_OF_FILE_CHECKPOINT, "compatibility");
|
|
|
|
|
|
|
|
#ifdef _WIN32
|
|
|
|
DWORD len;
|
2022-02-15 14:03:15 +01:00
|
|
|
ret= WriteFile(file, log_hdr_buf, sizeof log_hdr_buf,
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
&len, nullptr) && len == sizeof log_hdr_buf;
|
|
|
|
#else
|
2022-02-15 14:03:15 +01:00
|
|
|
ret= sizeof log_hdr_buf == write(file, log_hdr_buf, sizeof log_hdr_buf);
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
#endif
|
2022-02-15 14:03:15 +01:00
|
|
|
if (!os_file_close_func(file) || !ret)
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
goto invalid_log;
|
|
|
|
return false;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* ================= common ================= */
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Read backup meta info.
|
|
|
|
@return TRUE on success, FALSE on failure. */
|
|
|
|
static
|
|
|
|
my_bool
|
|
|
|
xtrabackup_read_metadata(char *filename)
|
|
|
|
{
|
|
|
|
FILE *fp;
|
|
|
|
my_bool r = TRUE;
|
|
|
|
|
|
|
|
fp = fopen(filename,"r");
|
|
|
|
if(!fp) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: cannot open %s", filename);
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (fscanf(fp, "backup_type = %29s\n", metadata_type)
|
|
|
|
!= 1) {
|
|
|
|
r = FALSE;
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
/* Use UINT64PF instead of LSN_PF here, as we have to maintain the file
|
|
|
|
format. */
|
|
|
|
if (fscanf(fp, "from_lsn = " UINT64PF "\n", &metadata_from_lsn)
|
|
|
|
!= 1) {
|
|
|
|
r = FALSE;
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
if (fscanf(fp, "to_lsn = " UINT64PF "\n", &metadata_to_lsn)
|
|
|
|
!= 1) {
|
|
|
|
r = FALSE;
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
if (fscanf(fp, "last_lsn = " UINT64PF "\n", &metadata_last_lsn)
|
|
|
|
!= 1) {
|
|
|
|
metadata_last_lsn = 0;
|
|
|
|
}
|
|
|
|
/* Optional fields */
|
|
|
|
|
|
|
|
end:
|
|
|
|
fclose(fp);
|
|
|
|
|
|
|
|
return(r);
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Print backup meta info to a specified buffer. */
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xtrabackup_print_metadata(char *buf, size_t buf_len)
|
|
|
|
{
|
|
|
|
/* Use UINT64PF instead of LSN_PF here, as we have to maintain the file
|
|
|
|
format. */
|
|
|
|
snprintf(buf, buf_len,
|
|
|
|
"backup_type = %s\n"
|
|
|
|
"from_lsn = " UINT64PF "\n"
|
|
|
|
"to_lsn = " UINT64PF "\n"
|
2019-03-14 16:59:27 +01:00
|
|
|
"last_lsn = " UINT64PF "\n",
|
2017-04-18 20:43:20 +02:00
|
|
|
metadata_type,
|
|
|
|
metadata_from_lsn,
|
|
|
|
metadata_to_lsn,
|
2019-03-14 16:59:27 +01:00
|
|
|
metadata_last_lsn);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Stream backup meta info to a specified datasink.
|
|
|
|
@return TRUE on success, FALSE on failure. */
|
|
|
|
static
|
|
|
|
my_bool
|
|
|
|
xtrabackup_stream_metadata(ds_ctxt_t *ds_ctxt)
|
|
|
|
{
|
|
|
|
char buf[1024];
|
|
|
|
size_t len;
|
|
|
|
ds_file_t *stream;
|
|
|
|
MY_STAT mystat;
|
|
|
|
my_bool rc = TRUE;
|
|
|
|
|
|
|
|
xtrabackup_print_metadata(buf, sizeof(buf));
|
|
|
|
|
|
|
|
len = strlen(buf);
|
|
|
|
|
|
|
|
mystat.st_size = len;
|
|
|
|
mystat.st_mtime = my_time(0);
|
|
|
|
|
|
|
|
stream = ds_open(ds_ctxt, XTRABACKUP_METADATA_FILENAME, &mystat);
|
|
|
|
if (stream == NULL) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: cannot open output stream for %s", XTRABACKUP_METADATA_FILENAME);
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (ds_write(stream, buf, len)) {
|
|
|
|
rc = FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (ds_close(stream)) {
|
|
|
|
rc = FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
return(rc);
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Write backup meta info to a specified file.
|
|
|
|
@return TRUE on success, FALSE on failure. */
|
|
|
|
static
|
|
|
|
my_bool
|
|
|
|
xtrabackup_write_metadata(const char *filepath)
|
|
|
|
{
|
|
|
|
char buf[1024];
|
|
|
|
size_t len;
|
|
|
|
FILE *fp;
|
|
|
|
|
|
|
|
xtrabackup_print_metadata(buf, sizeof(buf));
|
|
|
|
|
|
|
|
len = strlen(buf);
|
|
|
|
|
|
|
|
fp = fopen(filepath, "w");
|
|
|
|
if(!fp) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: cannot open %s", filepath);
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
if (fwrite(buf, len, 1, fp) < 1) {
|
|
|
|
fclose(fp);
|
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
fclose(fp);
|
|
|
|
|
|
|
|
return(TRUE);
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Read meta info for an incremental delta.
|
|
|
|
@return TRUE on success, FALSE on failure. */
|
|
|
|
static my_bool
|
|
|
|
xb_read_delta_metadata(const char *filepath, xb_delta_info_t *info)
|
|
|
|
{
|
|
|
|
FILE* fp;
|
|
|
|
char key[51];
|
|
|
|
char value[51];
|
|
|
|
my_bool r = TRUE;
|
|
|
|
|
|
|
|
/* set defaults */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
ulint page_size = ULINT_UNDEFINED, zip_size = 0;
|
2021-07-22 10:22:47 +02:00
|
|
|
info->space_id = UINT32_MAX;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
fp = fopen(filepath, "r");
|
|
|
|
if (!fp) {
|
|
|
|
/* Meta files for incremental deltas are optional */
|
|
|
|
return(TRUE);
|
|
|
|
}
|
|
|
|
|
|
|
|
while (!feof(fp)) {
|
|
|
|
if (fscanf(fp, "%50s = %50s\n", key, value) == 2) {
|
|
|
|
if (strcmp(key, "page_size") == 0) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
page_size = strtoul(value, NULL, 10);
|
2017-04-18 20:43:20 +02:00
|
|
|
} else if (strcmp(key, "zip_size") == 0) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
zip_size = strtoul(value, NULL, 10);
|
2017-04-18 20:43:20 +02:00
|
|
|
} else if (strcmp(key, "space_id") == 0) {
|
2021-07-22 10:22:47 +02:00
|
|
|
info->space_id = static_cast<uint32_t>
|
|
|
|
(strtoul(value, NULL, 10));
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
fclose(fp);
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (page_size == ULINT_UNDEFINED) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("page_size is required in %s", filepath);
|
2017-04-18 20:43:20 +02:00
|
|
|
r = FALSE;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
} else {
|
2022-09-14 10:39:30 +02:00
|
|
|
info->page_size = page_size;
|
|
|
|
info->zip_size = zip_size;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2021-07-22 10:22:47 +02:00
|
|
|
if (info->space_id == UINT32_MAX) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: Warning: This backup was taken with XtraBackup 2.0.1 "
|
2017-04-18 20:43:20 +02:00
|
|
|
"or earlier, some DDL operations between full and incremental "
|
2019-01-14 22:28:23 +01:00
|
|
|
"backups may be handled incorrectly");
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
return(r);
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Write meta info for an incremental delta.
|
|
|
|
@return TRUE on success, FALSE on failure. */
|
|
|
|
my_bool
|
|
|
|
xb_write_delta_metadata(const char *filename, const xb_delta_info_t *info)
|
|
|
|
{
|
|
|
|
ds_file_t *f;
|
|
|
|
char buf[64];
|
|
|
|
my_bool ret;
|
|
|
|
size_t len;
|
|
|
|
MY_STAT mystat;
|
|
|
|
|
|
|
|
snprintf(buf, sizeof(buf),
|
2017-09-19 19:45:17 +02:00
|
|
|
"page_size = " ULINTPF "\n"
|
|
|
|
"zip_size = " ULINTPF " \n"
|
2021-07-22 10:22:47 +02:00
|
|
|
"space_id = %u\n",
|
2019-02-06 18:50:11 +01:00
|
|
|
info->page_size,
|
|
|
|
info->zip_size,
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
info->space_id);
|
2017-04-18 20:43:20 +02:00
|
|
|
len = strlen(buf);
|
|
|
|
|
|
|
|
mystat.st_size = len;
|
|
|
|
mystat.st_mtime = my_time(0);
|
|
|
|
|
|
|
|
f = ds_open(ds_meta, filename, &mystat);
|
|
|
|
if (f == NULL) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: Can't open output stream for %s",filename);
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
ret = (ds_write(f, buf, len) == 0);
|
|
|
|
|
|
|
|
if (ds_close(f)) {
|
|
|
|
ret = FALSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
return(ret);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* ================= backup ================= */
|
2020-12-04 15:18:04 +01:00
|
|
|
void xtrabackup_io_throttling()
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2022-08-18 16:12:00 +02:00
|
|
|
if (!xtrabackup_backup || !xtrabackup_throttle)
|
2020-12-04 15:18:04 +01:00
|
|
|
return;
|
|
|
|
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2022-08-18 16:12:00 +02:00
|
|
|
if (io_ticket-- < 0)
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_cond_wait(&wait_throttle, &recv_sys.mutex);
|
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
static
|
|
|
|
my_bool regex_list_check_match(
|
|
|
|
const regex_list_t& list,
|
|
|
|
const char* name)
|
|
|
|
{
|
|
|
|
regmatch_t tables_regmatch[1];
|
|
|
|
for (regex_list_t::const_iterator i = list.begin(), end = list.end();
|
|
|
|
i != end; ++i) {
|
|
|
|
const regex_t& regex = *i;
|
|
|
|
int regres = regexec(®ex, name, 1, tables_regmatch, 0);
|
|
|
|
|
|
|
|
if (regres != REG_NOMATCH) {
|
|
|
|
return(TRUE);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
static
|
|
|
|
my_bool
|
|
|
|
find_filter_in_hashtable(
|
|
|
|
const char* name,
|
|
|
|
hash_table_t* table,
|
|
|
|
xb_filter_entry_t** result
|
|
|
|
)
|
|
|
|
{
|
|
|
|
xb_filter_entry_t* found = NULL;
|
2021-06-14 08:15:20 +02:00
|
|
|
const ulint fold = my_crc32c(0, name, strlen(name));
|
|
|
|
HASH_SEARCH(name_hash, table, fold,
|
2017-04-19 15:09:03 +02:00
|
|
|
xb_filter_entry_t*,
|
|
|
|
found, (void) 0,
|
|
|
|
!strcmp(found->name, name));
|
|
|
|
|
|
|
|
if (found && result) {
|
|
|
|
*result = found;
|
|
|
|
}
|
|
|
|
return (found != NULL);
|
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/************************************************************************
|
2017-04-19 15:09:03 +02:00
|
|
|
Checks if a given table name matches any of specifications given in
|
|
|
|
regex_list or tables_hash.
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
@return TRUE on match or both regex_list and tables_hash are empty.*/
|
2017-04-18 20:43:20 +02:00
|
|
|
static my_bool
|
2017-04-19 15:09:03 +02:00
|
|
|
check_if_table_matches_filters(const char *name,
|
|
|
|
const regex_list_t& regex_list,
|
|
|
|
hash_table_t* tables_hash)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2020-06-18 11:26:28 +02:00
|
|
|
if (regex_list.empty() && !tables_hash->array) {
|
2017-04-19 15:09:03 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
if (regex_list_check_match(regex_list, name)) {
|
|
|
|
return(TRUE);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
return tables_hash->array &&
|
|
|
|
find_filter_in_hashtable(name, tables_hash, NULL);
|
2017-04-19 15:09:03 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
enum skip_database_check_result {
|
|
|
|
DATABASE_SKIP,
|
|
|
|
DATABASE_SKIP_SOME_TABLES,
|
|
|
|
DATABASE_DONT_SKIP,
|
|
|
|
DATABASE_DONT_SKIP_UNLESS_EXPLICITLY_EXCLUDED,
|
|
|
|
};
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
/************************************************************************
|
|
|
|
Checks if a database specified by name should be skipped from backup based on
|
|
|
|
the --databases, --databases_file or --databases_exclude options.
|
|
|
|
|
|
|
|
@return TRUE if entire database should be skipped,
|
|
|
|
FALSE otherwise.
|
|
|
|
*/
|
|
|
|
static
|
|
|
|
skip_database_check_result
|
|
|
|
check_if_skip_database(
|
|
|
|
const char* name /*!< in: path to the database */
|
|
|
|
)
|
|
|
|
{
|
|
|
|
/* There are some filters for databases, check them */
|
|
|
|
xb_filter_entry_t* database = NULL;
|
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
if (databases_exclude_hash.array &&
|
|
|
|
find_filter_in_hashtable(name, &databases_exclude_hash,
|
2017-04-19 15:09:03 +02:00
|
|
|
&database) &&
|
2020-10-22 16:08:49 +02:00
|
|
|
(!database->has_tables || !databases_include_hash.array)) {
|
2017-04-19 15:09:03 +02:00
|
|
|
/* Database is found and there are no tables specified,
|
|
|
|
skip entire db. */
|
|
|
|
return DATABASE_SKIP;
|
|
|
|
}
|
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
if (databases_include_hash.array) {
|
|
|
|
if (!find_filter_in_hashtable(name, &databases_include_hash,
|
2017-04-19 15:09:03 +02:00
|
|
|
&database)) {
|
|
|
|
/* Database isn't found, skip the database */
|
|
|
|
return DATABASE_SKIP;
|
|
|
|
} else if (database->has_tables) {
|
|
|
|
return DATABASE_SKIP_SOME_TABLES;
|
|
|
|
} else {
|
|
|
|
return DATABASE_DONT_SKIP_UNLESS_EXPLICITLY_EXCLUDED;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
return DATABASE_DONT_SKIP;
|
|
|
|
}
|
|
|
|
|
|
|
|
/************************************************************************
|
|
|
|
Checks if a database specified by path should be skipped from backup based on
|
|
|
|
the --databases, --databases_file or --databases_exclude options.
|
|
|
|
|
|
|
|
@return TRUE if the table should be skipped. */
|
|
|
|
my_bool
|
|
|
|
check_if_skip_database_by_path(
|
|
|
|
const char* path /*!< in: path to the db directory. */
|
|
|
|
)
|
|
|
|
{
|
2020-06-18 11:26:28 +02:00
|
|
|
if (!databases_include_hash.array && !databases_exclude_hash.array) {
|
2017-04-19 15:09:03 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
const char* db_name = strrchr(path, '/');
|
|
|
|
#ifdef _WIN32
|
|
|
|
if (const char* last = strrchr(path, '\\')) {
|
|
|
|
if (!db_name || last > db_name) {
|
|
|
|
db_name = last;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
if (db_name == NULL) {
|
|
|
|
db_name = path;
|
|
|
|
} else {
|
|
|
|
++db_name;
|
|
|
|
}
|
|
|
|
|
|
|
|
return check_if_skip_database(db_name) == DATABASE_SKIP;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/************************************************************************
|
|
|
|
Checks if a table specified as a name in the form "database/name" (InnoDB 5.6)
|
|
|
|
or "./database/name.ibd" (InnoDB 5.5-) should be skipped from backup based on
|
|
|
|
the --tables or --tables-file options.
|
|
|
|
|
|
|
|
@return TRUE if the table should be skipped. */
|
|
|
|
my_bool
|
|
|
|
check_if_skip_table(
|
|
|
|
/******************/
|
|
|
|
const char* name) /*!< in: path to the table */
|
|
|
|
{
|
|
|
|
char buf[FN_REFLEN];
|
|
|
|
const char *dbname, *tbname;
|
|
|
|
const char *ptr;
|
|
|
|
char *eptr;
|
|
|
|
|
2019-01-14 14:25:08 +01:00
|
|
|
|
|
|
|
dbname = NULL;
|
|
|
|
tbname = name;
|
2021-06-17 12:46:16 +02:00
|
|
|
for (;;) {
|
|
|
|
ptr= strchr(tbname, '/');
|
|
|
|
#ifdef _WIN32
|
|
|
|
if (!ptr) {
|
|
|
|
ptr= strchr(tbname,'\\');
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
if (!ptr) {
|
|
|
|
break;
|
|
|
|
}
|
2019-01-14 14:25:08 +01:00
|
|
|
dbname = tbname;
|
|
|
|
tbname = ptr + 1;
|
|
|
|
}
|
|
|
|
|
2021-06-17 12:46:16 +02:00
|
|
|
if (strncmp(tbname, tmp_file_prefix, tmp_file_prefix_length) == 0) {
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
if (regex_exclude_list.empty() &&
|
|
|
|
regex_include_list.empty() &&
|
2020-06-18 11:26:28 +02:00
|
|
|
!tables_include_hash.array &&
|
|
|
|
!tables_exclude_hash.array &&
|
|
|
|
!databases_include_hash.array &&
|
|
|
|
!databases_exclude_hash.array) {
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (dbname == NULL) {
|
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
2019-03-06 10:22:27 +01:00
|
|
|
strncpy(buf, dbname, FN_REFLEN - 1);
|
|
|
|
buf[FN_REFLEN - 1] = '\0';
|
|
|
|
buf[tbname - 1 - dbname] = '\0';
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
const skip_database_check_result skip_database =
|
|
|
|
check_if_skip_database(buf);
|
|
|
|
if (skip_database == DATABASE_SKIP) {
|
|
|
|
return (TRUE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
buf[tbname - 1 - dbname] = '.';
|
|
|
|
|
|
|
|
/* Check if there's a suffix in the table name. If so, truncate it. We
|
|
|
|
rely on the fact that a dot cannot be a part of a table name (it is
|
|
|
|
encoded by the server with the @NNNN syntax). */
|
|
|
|
if ((eptr = strchr(&buf[tbname - dbname], '.')) != NULL) {
|
|
|
|
|
|
|
|
*eptr = '\0';
|
|
|
|
}
|
|
|
|
|
|
|
|
/* For partitioned tables first try to match against the regexp
|
|
|
|
without truncating the #P#... suffix so we can backup individual
|
|
|
|
partitions with regexps like '^test[.]t#P#p5' */
|
2017-04-19 15:09:03 +02:00
|
|
|
if (check_if_table_matches_filters(buf, regex_exclude_list,
|
2020-06-18 11:26:28 +02:00
|
|
|
&tables_exclude_hash)) {
|
2017-04-19 15:09:03 +02:00
|
|
|
return(TRUE);
|
|
|
|
}
|
|
|
|
if (check_if_table_matches_filters(buf, regex_include_list,
|
2020-06-18 11:26:28 +02:00
|
|
|
&tables_include_hash)) {
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
if ((eptr = strstr(buf, "#P#")) != NULL) {
|
|
|
|
*eptr = 0;
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
if (check_if_table_matches_filters(buf, regex_exclude_list,
|
2020-06-18 11:26:28 +02:00
|
|
|
&tables_exclude_hash)) {
|
2017-04-19 15:09:03 +02:00
|
|
|
return (TRUE);
|
|
|
|
}
|
|
|
|
if (check_if_table_matches_filters(buf, regex_include_list,
|
2020-06-18 11:26:28 +02:00
|
|
|
&tables_include_hash)) {
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
if (skip_database == DATABASE_DONT_SKIP_UNLESS_EXPLICITLY_EXCLUDED) {
|
|
|
|
/* Database is in include-list, and qualified name wasn't
|
|
|
|
found in any of exclusion filters.*/
|
|
|
|
return (FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (skip_database == DATABASE_SKIP_SOME_TABLES ||
|
|
|
|
!regex_include_list.empty() ||
|
2020-06-18 11:26:28 +02:00
|
|
|
tables_include_hash.array) {
|
2017-04-19 15:09:03 +02:00
|
|
|
|
|
|
|
/* Include lists are present, but qualified name
|
|
|
|
failed to match any.*/
|
|
|
|
return(TRUE);
|
|
|
|
}
|
|
|
|
|
|
|
|
return(FALSE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
const char*
|
|
|
|
xb_get_copy_action(const char *dflt)
|
|
|
|
{
|
|
|
|
const char *action;
|
|
|
|
|
|
|
|
if (xtrabackup_stream) {
|
|
|
|
if (xtrabackup_compress) {
|
2017-05-26 11:13:48 +02:00
|
|
|
action = "Compressing and streaming";
|
2017-04-18 20:43:20 +02:00
|
|
|
} else {
|
|
|
|
action = "Streaming";
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
if (xtrabackup_compress) {
|
2017-05-26 11:13:48 +02:00
|
|
|
action = "Compressing";
|
2017-04-18 20:43:20 +02:00
|
|
|
} else {
|
|
|
|
action = dflt;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return(action);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
/** Copy innodb data file to the specified destination.
|
|
|
|
|
|
|
|
@param[in] node file node of a tablespace
|
|
|
|
@param[in] thread_n thread id, used in the text of diagnostic messages
|
|
|
|
@param[in] dest_name destination file name
|
|
|
|
@param[in] write_filter write filter to copy data, can be pass-through filter
|
|
|
|
for full backup, pages filter for incremental backup, etc.
|
|
|
|
|
|
|
|
@return FALSE on success and TRUE on error */
|
|
|
|
static my_bool xtrabackup_copy_datafile(fil_node_t *node, uint thread_n,
|
|
|
|
const char *dest_name,
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
const xb_write_filt_t &write_filter,
|
|
|
|
CorruptedPages &corrupted_pages)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
|
|
|
char dst_name[FN_REFLEN];
|
|
|
|
ds_file_t *dstfile = NULL;
|
|
|
|
xb_fil_cur_t cursor;
|
|
|
|
xb_fil_cur_result_t res;
|
|
|
|
xb_write_filt_ctxt_t write_filt_ctxt;
|
|
|
|
const char *action;
|
|
|
|
xb_read_filt_t *read_filter;
|
|
|
|
my_bool rc = FALSE;
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (fil_is_user_tablespace_id(node->space->id)
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
&& check_if_skip_table(filename_to_spacename(node->name,
|
|
|
|
strlen(node->name)).
|
|
|
|
c_str())) {
|
|
|
|
msg(thread_n, "Skipping %s.", node->name);
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
2020-10-27 18:03:41 +01:00
|
|
|
memset(&write_filt_ctxt, 0, sizeof(xb_write_filt_ctxt_t));
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
bool was_dropped;
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2018-08-09 16:06:52 +02:00
|
|
|
was_dropped = (ddl_tracker.drops.find(node->space->id) != ddl_tracker.drops.end());
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2018-08-09 16:06:52 +02:00
|
|
|
if (was_dropped) {
|
2018-08-16 07:54:58 +02:00
|
|
|
if (node->is_open()) {
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_lock(&fil_system.mutex);
|
2018-08-16 07:54:58 +02:00
|
|
|
node->close();
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_unlock(&fil_system.mutex);
|
2018-08-16 07:54:58 +02:00
|
|
|
}
|
2018-08-09 16:06:52 +02:00
|
|
|
goto skip;
|
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
if (!changed_page_bitmap) {
|
|
|
|
read_filter = &rf_pass_through;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
read_filter = &rf_bitmap;
|
|
|
|
}
|
2018-08-09 16:06:52 +02:00
|
|
|
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
res = xb_fil_cur_open(&cursor, read_filter, node, thread_n, ULLONG_MAX);
|
2017-04-18 20:43:20 +02:00
|
|
|
if (res == XB_FIL_CUR_SKIP) {
|
|
|
|
goto skip;
|
|
|
|
} else if (res == XB_FIL_CUR_ERROR) {
|
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
2019-03-21 07:54:35 +01:00
|
|
|
strncpy(dst_name, dest_name ? dest_name : cursor.rel_path,
|
|
|
|
sizeof dst_name - 1);
|
|
|
|
dst_name[sizeof dst_name - 1] = '\0';
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
ut_a(write_filter.process != NULL);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
if (write_filter.init != NULL &&
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
!write_filter.init(&write_filt_ctxt, dst_name, &cursor,
|
|
|
|
opt_log_innodb_page_corruption ? &corrupted_pages : NULL)) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg (thread_n, "mariabackup: error: failed to initialize page write filter.");
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
|
|
|
dstfile = ds_open(ds_data, dst_name, &cursor.statinfo);
|
|
|
|
if (dstfile == NULL) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg(thread_n,"mariabackup: error: can't open the destination stream for %s", dst_name);
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
|
|
|
action = xb_get_copy_action();
|
|
|
|
|
|
|
|
if (xtrabackup_stream) {
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
msg(thread_n, "%s %s", action, node->name);
|
2017-04-18 20:43:20 +02:00
|
|
|
} else {
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
msg(thread_n, "%s %s to %s", action, node->name,
|
|
|
|
dstfile->path);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* The main copy loop */
|
2023-01-02 08:43:59 +01:00
|
|
|
while (1) {
|
|
|
|
res = xb_fil_cur_read(&cursor, corrupted_pages);
|
|
|
|
if (res == XB_FIL_CUR_ERROR) {
|
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (res == XB_FIL_CUR_EOF) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
if (!write_filter.process(&write_filt_ctxt, dstfile)) {
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
2023-01-02 08:43:59 +01:00
|
|
|
if (res == XB_FIL_CUR_SKIP) {
|
2023-01-10 13:42:50 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2023-01-02 08:43:59 +01:00
|
|
|
fail_undo_ids.insert(
|
|
|
|
static_cast<uint32_t>(cursor.space_id));
|
2023-01-10 13:42:50 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2023-01-02 08:43:59 +01:00
|
|
|
break;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
if (write_filter.finalize
|
|
|
|
&& !write_filter.finalize(&write_filt_ctxt, dstfile)) {
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
} else {
|
|
|
|
const fil_space_t::name_type name = node->space->name();
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
ddl_tracker.tables_in_backup.emplace(node->space->id,
|
|
|
|
std::string(name.data(),
|
|
|
|
name.size()));
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
}
|
2018-08-09 16:06:52 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* close */
|
2019-01-14 22:28:23 +01:00
|
|
|
msg(thread_n," ...done");
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_fil_cur_close(&cursor);
|
|
|
|
if (ds_close(dstfile)) {
|
|
|
|
rc = TRUE;
|
|
|
|
}
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
if (write_filter.deinit) {
|
|
|
|
write_filter.deinit(&write_filt_ctxt);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
return(rc);
|
|
|
|
|
|
|
|
error:
|
|
|
|
xb_fil_cur_close(&cursor);
|
|
|
|
if (dstfile != NULL) {
|
|
|
|
ds_close(dstfile);
|
|
|
|
}
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
if (write_filter.deinit) {
|
|
|
|
write_filter.deinit(&write_filt_ctxt);;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2019-01-14 22:28:23 +01:00
|
|
|
msg(thread_n, "mariabackup: xtrabackup_copy_datafile() failed.");
|
2017-04-18 20:43:20 +02:00
|
|
|
return(TRUE); /*ERROR*/
|
|
|
|
|
|
|
|
skip:
|
|
|
|
|
|
|
|
if (dstfile != NULL) {
|
|
|
|
ds_close(dstfile);
|
|
|
|
}
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
if (write_filter.deinit) {
|
|
|
|
write_filter.deinit(&write_filt_ctxt);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
msg(thread_n,"Warning: We assume the table was dropped during xtrabackup execution and ignore the tablespace %s", node->name);
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
/** Copy redo log until the current end of the log is reached
|
|
|
|
@return whether the operation failed */
|
|
|
|
static bool xtrabackup_copy_logfile()
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_assert_owner(&recv_sys.mutex);
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
DBUG_EXECUTE_IF("log_checksum_mismatch", return false;);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
ut_a(dst_log_file);
|
|
|
|
ut_ad(recv_sys.is_initialised());
|
|
|
|
const size_t sequence_offset{log_sys.is_encrypted() ? 8U + 5U : 5U};
|
|
|
|
const size_t block_size_1{log_sys.get_block_size() - 1};
|
2018-03-13 10:49:30 +01:00
|
|
|
|
2022-07-01 17:07:07 +02:00
|
|
|
ut_ad(!log_sys.is_pmem());
|
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
{
|
|
|
|
recv_sys.offset= size_t(recv_sys.lsn - log_sys.get_first_lsn()) &
|
|
|
|
block_size_1;
|
|
|
|
recv_sys.len= 0;
|
|
|
|
}
|
2017-04-18 21:05:57 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
for (unsigned retry_count{0};;)
|
|
|
|
{
|
|
|
|
recv_sys_t::parse_mtr_result r;
|
|
|
|
size_t start_offset{recv_sys.offset};
|
2017-04-18 21:05:57 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
{
|
|
|
|
{
|
|
|
|
auto source_offset=
|
|
|
|
log_sys.calc_lsn_offset(recv_sys.lsn + recv_sys.len -
|
|
|
|
recv_sys.offset);
|
|
|
|
source_offset&= ~block_size_1;
|
|
|
|
size_t size{log_sys.buf_size - recv_sys.len};
|
2022-01-27 15:17:40 +01:00
|
|
|
if (UNIV_UNLIKELY(source_offset + size > log_sys.file_size))
|
|
|
|
{
|
|
|
|
const size_t first{size_t(log_sys.file_size - source_offset)};
|
|
|
|
ut_ad(first <= log_sys.buf_size);
|
|
|
|
log_sys.log.read(source_offset, {log_sys.buf, first});
|
|
|
|
size-= first;
|
|
|
|
if (log_sys.START_OFFSET + size > source_offset)
|
|
|
|
size= size_t(source_offset - log_sys.START_OFFSET);
|
|
|
|
if (size)
|
|
|
|
log_sys.log.read(log_sys.START_OFFSET,
|
|
|
|
{log_sys.buf + first, size});
|
|
|
|
size+= first;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
log_sys.log.read(source_offset, {log_sys.buf, size});
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
recv_sys.len= size;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (log_sys.buf[recv_sys.offset] <= 1)
|
|
|
|
break;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2022-02-10 14:39:27 +01:00
|
|
|
if (recv_sys.parse_mtr(STORE_NO) == recv_sys_t::OK)
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
{
|
|
|
|
do
|
|
|
|
{
|
|
|
|
/* Set the sequence bit (the backed-up log will not wrap around) */
|
|
|
|
byte *seq= &log_sys.buf[recv_sys.offset - sequence_offset];
|
|
|
|
ut_ad(*seq == log_sys.get_sequence_bit(recv_sys.lsn -
|
|
|
|
sequence_offset));
|
|
|
|
*seq= 1;
|
|
|
|
}
|
2022-02-10 14:39:27 +01:00
|
|
|
while ((r= recv_sys.parse_mtr(STORE_NO)) == recv_sys_t::OK);
|
2018-03-13 10:49:30 +01:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (ds_write(dst_log_file, log_sys.buf + start_offset,
|
|
|
|
recv_sys.offset - start_offset))
|
|
|
|
{
|
|
|
|
msg("Error: write to ib_logfile0 failed");
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
const auto ofs= recv_sys.offset & ~block_size_1;
|
|
|
|
memmove_aligned<64>(log_sys.buf, log_sys.buf + ofs,
|
|
|
|
recv_sys.len - ofs);
|
|
|
|
recv_sys.len-= ofs;
|
|
|
|
recv_sys.offset&= block_size_1;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
pthread_cond_broadcast(&scanned_lsn_cond);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (r == recv_sys_t::GOT_EOF)
|
|
|
|
break;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (recv_sys.offset < log_sys.get_block_size())
|
|
|
|
break;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (xtrabackup_throttle && io_ticket-- < 0)
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_cond_wait(&wait_throttle, &recv_sys.mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
retry_count= 0;
|
2022-02-10 15:37:12 +01:00
|
|
|
continue;
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
recv_sys.len= recv_sys.offset & ~block_size_1;
|
|
|
|
if (retry_count == 100)
|
|
|
|
break;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
|
|
|
if (!retry_count++)
|
|
|
|
msg("Retrying read of log at LSN=" LSN_PF, recv_sys.lsn);
|
|
|
|
my_sleep(1000);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
msg(">> log scanned up to (" LSN_PF ")", recv_sys.lsn);
|
|
|
|
return false;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
/**
|
|
|
|
Wait until redo log copying thread processes given lsn
|
|
|
|
*/
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
void backup_wait_for_lsn(lsn_t lsn)
|
|
|
|
{
|
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
|
|
|
for (lsn_t last_lsn{recv_sys.lsn}; last_lsn < lsn; )
|
|
|
|
{
|
|
|
|
timespec abstime;
|
|
|
|
set_timespec(abstime, 5);
|
|
|
|
if (my_cond_timedwait(&scanned_lsn_cond, &recv_sys.mutex.m_mutex,
|
|
|
|
&abstime) &&
|
|
|
|
last_lsn == recv_sys.lsn)
|
|
|
|
die("Was only able to copy log from " LSN_PF " to " LSN_PF
|
|
|
|
", not " LSN_PF "; try increasing innodb_log_file_size",
|
|
|
|
log_sys.next_checkpoint_lsn, last_lsn, lsn);
|
|
|
|
last_lsn= recv_sys.lsn;
|
|
|
|
}
|
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
extern lsn_t server_lsn_after_lock;
|
|
|
|
|
2021-03-17 08:03:06 +01:00
|
|
|
static void log_copying_thread()
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2020-12-04 15:18:04 +01:00
|
|
|
my_thread_init();
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2020-12-04 15:18:04 +01:00
|
|
|
while (!xtrabackup_copy_logfile() &&
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
(!metadata_to_lsn || metadata_to_lsn > recv_sys.lsn))
|
2020-12-04 15:18:04 +01:00
|
|
|
{
|
|
|
|
timespec abstime;
|
2022-04-15 17:21:54 +02:00
|
|
|
set_timespec_nsec(abstime, 1000000ULL * xtrabackup_log_copy_interval);
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_cond_timedwait(&log_copying_stop, &recv_sys.mutex, &abstime);
|
2020-12-04 15:18:04 +01:00
|
|
|
}
|
|
|
|
log_copying_running= false;
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2020-12-04 15:18:04 +01:00
|
|
|
my_thread_end();
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2022-12-21 12:41:10 +01:00
|
|
|
/** whether io_watching_thread() is active; protected by log_sys.mutex */
|
2020-12-04 15:18:04 +01:00
|
|
|
static bool have_io_watching_thread;
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* io throttle watching (rough) */
|
2022-12-21 12:41:10 +01:00
|
|
|
static void io_watching_thread()
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2022-12-21 12:41:10 +01:00
|
|
|
my_thread_init();
|
2020-12-04 15:18:04 +01:00
|
|
|
/* currently, for --backup only */
|
2022-12-21 12:41:10 +01:00
|
|
|
ut_ad(xtrabackup_backup);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2022-12-21 12:41:10 +01:00
|
|
|
ut_ad(have_io_watching_thread);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-12-04 15:18:04 +01:00
|
|
|
while (log_copying_running && !metadata_to_lsn)
|
|
|
|
{
|
|
|
|
timespec abstime;
|
|
|
|
set_timespec(abstime, 1);
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_cond_timedwait(&log_copying_stop, &recv_sys.mutex, &abstime);
|
2020-12-04 15:18:04 +01:00
|
|
|
io_ticket= xtrabackup_throttle;
|
|
|
|
mysql_cond_broadcast(&wait_throttle);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-12-04 15:18:04 +01:00
|
|
|
/* stop io throttle */
|
|
|
|
xtrabackup_throttle= 0;
|
2022-12-21 12:41:10 +01:00
|
|
|
have_io_watching_thread= false;
|
2020-12-04 15:18:04 +01:00
|
|
|
mysql_cond_broadcast(&wait_throttle);
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2022-12-21 12:41:10 +01:00
|
|
|
my_thread_end();
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
#ifndef DBUG_OFF
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
char *dbug_mariabackup_get_val(const char *event,
|
|
|
|
const fil_space_t::name_type key)
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
{
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
char envvar[FN_REFLEN];
|
|
|
|
strncpy(envvar, event, sizeof envvar - 1);
|
|
|
|
envvar[(sizeof envvar) - 1] = '\0';
|
|
|
|
|
|
|
|
if (key.size() && key.size() + strlen(envvar) < (sizeof envvar) - 2)
|
|
|
|
{
|
|
|
|
strcat(envvar, "_");
|
|
|
|
strncat(envvar, key.data(), key.size());
|
|
|
|
if (char *slash= strchr(envvar, '/'))
|
|
|
|
*slash= '_';
|
|
|
|
}
|
|
|
|
|
|
|
|
char *val = getenv(envvar);
|
|
|
|
return val && *val ? val : nullptr;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
}
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
/*
|
|
|
|
In debug mode, execute SQL statement that was passed via environment.
|
|
|
|
To use this facility, you need to
|
|
|
|
|
|
|
|
1. Add code DBUG_EXECUTE_MARIABACKUP_EVENT("my_event_name", key););
|
|
|
|
to the code. key is usually a table name
|
|
|
|
2. Set environment variable my_event_name_$key SQL statement you want to execute
|
|
|
|
when event occurs, in DBUG_EXECUTE_IF from above.
|
|
|
|
In mtr , you can set environment via 'let' statement (do not use $ as the first char
|
|
|
|
for the variable)
|
|
|
|
3. start mariabackup with --dbug=+d,debug_mariabackup_events
|
|
|
|
*/
|
2021-06-17 12:46:16 +02:00
|
|
|
void dbug_mariabackup_event(const char *event,
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
const fil_space_t::name_type key)
|
2018-08-09 16:06:52 +02:00
|
|
|
{
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
char *sql = dbug_mariabackup_get_val(event, key);
|
|
|
|
if (sql && *sql) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("dbug_mariabackup_event : executing '%s'", sql);
|
2018-08-09 16:06:52 +02:00
|
|
|
xb_mysql_query(mysql_connection, sql, false, true);
|
|
|
|
}
|
|
|
|
}
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
#endif // DBUG_OFF
|
2018-08-09 16:06:52 +02:00
|
|
|
|
2021-03-17 08:03:06 +01:00
|
|
|
/** Datafiles copying thread.*/
|
|
|
|
static void data_copy_thread_func(data_thread_ctxt_t *ctxt) /* thread context */
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
|
|
|
uint num = ctxt->num;
|
|
|
|
fil_node_t* node;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
ut_ad(ctxt->corrupted_pages);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
Initialize mysys thread-specific memory so we can
|
|
|
|
use mysys functions in this thread.
|
|
|
|
*/
|
|
|
|
my_thread_init();
|
|
|
|
|
|
|
|
while ((node = datafiles_iter_next(ctxt->it)) != NULL) {
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
DBUG_MARIABACKUP_EVENT("before_copy", node->space->name());
|
|
|
|
DBUG_EXECUTE_FOR_KEY("wait_innodb_redo_before_copy",
|
|
|
|
node->space->name(),
|
2020-10-27 18:03:41 +01:00
|
|
|
backup_wait_for_lsn(get_current_lsn(mysql_connection)););
|
2017-04-18 20:43:20 +02:00
|
|
|
/* copy the datafile */
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
if (xtrabackup_copy_datafile(node, num, NULL,
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
xtrabackup_incremental ? wf_incremental : wf_write_through,
|
|
|
|
*ctxt->corrupted_pages))
|
2019-01-15 22:47:54 +01:00
|
|
|
die("failed to copy datafile.");
|
2018-08-09 16:06:52 +02:00
|
|
|
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
DBUG_MARIABACKUP_EVENT("after_copy", node->space->name());
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2018-09-20 20:27:59 +02:00
|
|
|
pthread_mutex_lock(ctxt->count_mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
(*ctxt->count)--;
|
2018-09-20 20:27:59 +02:00
|
|
|
pthread_mutex_unlock(ctxt->count_mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
my_thread_end();
|
|
|
|
}
|
|
|
|
|
|
|
|
/************************************************************************
|
|
|
|
Initialize the appropriate datasink(s). Both local backups and streaming in the
|
|
|
|
'xbstream' format allow parallel writes so we can write directly.
|
|
|
|
|
|
|
|
Otherwise (i.e. when streaming in the 'tar' format) we need 2 separate datasinks
|
|
|
|
for the data stream (and don't allow parallel data copying) and for metainfo
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
files (including ib_logfile0). The second datasink writes to temporary
|
2017-04-18 20:43:20 +02:00
|
|
|
files first, and then streams them in a serialized way when closed. */
|
|
|
|
static void
|
|
|
|
xtrabackup_init_datasinks(void)
|
|
|
|
{
|
|
|
|
/* Start building out the pipelines from the terminus back */
|
|
|
|
if (xtrabackup_stream) {
|
|
|
|
/* All streaming goes to stdout */
|
|
|
|
ds_data = ds_meta = ds_redo = ds_create(xtrabackup_target_dir,
|
|
|
|
DS_TYPE_STDOUT);
|
|
|
|
} else {
|
|
|
|
/* Local filesystem */
|
|
|
|
ds_data = ds_meta = ds_redo = ds_create(xtrabackup_target_dir,
|
|
|
|
DS_TYPE_LOCAL);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Track it for destruction */
|
|
|
|
xtrabackup_add_datasink(ds_data);
|
|
|
|
|
|
|
|
/* Stream formatting */
|
|
|
|
if (xtrabackup_stream) {
|
|
|
|
ds_ctxt_t *ds;
|
2017-05-22 00:19:06 +02:00
|
|
|
|
|
|
|
ut_a(xtrabackup_stream_fmt == XB_STREAM_FMT_XBSTREAM);
|
|
|
|
ds = ds_create(xtrabackup_target_dir, DS_TYPE_XBSTREAM);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
xtrabackup_add_datasink(ds);
|
|
|
|
|
|
|
|
ds_set_pipe(ds, ds_data);
|
|
|
|
ds_data = ds;
|
|
|
|
|
|
|
|
|
2017-05-22 00:19:06 +02:00
|
|
|
ds_redo = ds_meta = ds_data;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Compression for ds_data and ds_redo */
|
|
|
|
if (xtrabackup_compress) {
|
|
|
|
ds_ctxt_t *ds;
|
|
|
|
|
|
|
|
/* Use a 1 MB buffer for compressed output stream */
|
|
|
|
ds = ds_create(xtrabackup_target_dir, DS_TYPE_BUFFER);
|
|
|
|
ds_buffer_set_size(ds, 1024 * 1024);
|
|
|
|
xtrabackup_add_datasink(ds);
|
|
|
|
ds_set_pipe(ds, ds_data);
|
|
|
|
if (ds_data != ds_redo) {
|
|
|
|
ds_data = ds;
|
|
|
|
ds = ds_create(xtrabackup_target_dir, DS_TYPE_BUFFER);
|
|
|
|
ds_buffer_set_size(ds, 1024 * 1024);
|
|
|
|
xtrabackup_add_datasink(ds);
|
|
|
|
ds_set_pipe(ds, ds_redo);
|
|
|
|
ds_redo = ds;
|
|
|
|
} else {
|
|
|
|
ds_redo = ds_data = ds;
|
|
|
|
}
|
|
|
|
|
|
|
|
ds = ds_create(xtrabackup_target_dir, DS_TYPE_COMPRESS);
|
|
|
|
xtrabackup_add_datasink(ds);
|
|
|
|
ds_set_pipe(ds, ds_data);
|
|
|
|
if (ds_data != ds_redo) {
|
|
|
|
ds_data = ds;
|
|
|
|
ds = ds_create(xtrabackup_target_dir, DS_TYPE_COMPRESS);
|
|
|
|
xtrabackup_add_datasink(ds);
|
|
|
|
ds_set_pipe(ds, ds_redo);
|
|
|
|
ds_redo = ds;
|
|
|
|
} else {
|
|
|
|
ds_redo = ds_data = ds;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/************************************************************************
|
|
|
|
Destroy datasinks.
|
|
|
|
|
|
|
|
Destruction is done in the specific order to not violate their order in the
|
|
|
|
pipeline so that each datasink is able to flush data down the pipeline. */
|
|
|
|
static void xtrabackup_destroy_datasinks(void)
|
|
|
|
{
|
|
|
|
for (uint i = actual_datasinks; i > 0; i--) {
|
|
|
|
ds_destroy(datasinks[i-1]);
|
|
|
|
datasinks[i-1] = NULL;
|
|
|
|
}
|
|
|
|
ds_data = NULL;
|
|
|
|
ds_meta = NULL;
|
|
|
|
ds_redo = NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
#define SRV_MAX_N_PENDING_SYNC_IOS 100
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/** Initialize the tablespace cache subsystem. */
|
2017-04-18 20:43:20 +02:00
|
|
|
static
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
void
|
|
|
|
xb_fil_io_init()
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2018-03-28 08:00:06 +02:00
|
|
|
fil_system.create(srv_file_per_table ? 50000 : 5000);
|
2020-10-26 14:59:30 +01:00
|
|
|
fil_system.freeze_space_list = 1;
|
MDEV-23855: Improve InnoDB log checkpoint performance
After MDEV-15053, MDEV-22871, MDEV-23399 shifted the scalability
bottleneck, log checkpoints became a new bottleneck.
If innodb_io_capacity is set low or innodb_max_dirty_pct_lwm is
set high and the workload fits in the buffer pool, the page cleaner
thread will perform very little flushing. When we reach the capacity
of the circular redo log file ib_logfile0 and must initiate a checkpoint,
some 'furious flushing' will be necessary. (If innodb_flush_sync=OFF,
then flushing would continue at the innodb_io_capacity rate, and
writers would be throttled.)
We have the best chance of advancing the checkpoint LSN immediately
after a page flush batch has been completed. Hence, it is best to
perform checkpoints after every batch in the page cleaner thread,
attempting to run once per second.
By initiating high-priority flushing in the page cleaner as early
as possible, we aim to make the throughput more stable.
The function buf_flush_wait_flushed() used to sleep for 10ms, hoping
that the page cleaner thread would do something during that time.
The observed end result was that a large number of threads that call
log_free_check() would end up sleeping while nothing useful is happening.
We will revise the design so that in the default innodb_flush_sync=ON
mode, buf_flush_wait_flushed() will wake up the page cleaner thread
to perform the necessary flushing, and it will wait for a signal from
the page cleaner thread.
If innodb_io_capacity is set to a low value (causing the page cleaner to
throttle its work), a write workload would initially perform well, until
the capacity of the circular ib_logfile0 is reached and log_free_check()
will trigger checkpoints. At that point, the extra waiting in
buf_flush_wait_flushed() will start reducing throughput.
The page cleaner thread will also initiate log checkpoints after each
buf_flush_lists() call, because that is the best point of time for
the checkpoint LSN to advance by the maximum amount.
Even in 'furious flushing' mode we invoke buf_flush_lists() with
innodb_io_capacity_max pages at a time, and at the start of each
batch (in the log_flush() callback function that runs in a separate
task) we will invoke os_aio_wait_until_no_pending_writes(). This
tweak allows the checkpoint to advance in smaller steps and
significantly reduces the maximum latency. On an Intel Optane 960
NVMe SSD on Linux, it reduced from 4.6 seconds to 74 milliseconds.
On Microsoft Windows with a slower SSD, it reduced from more than
180 seconds to 0.6 seconds.
We will make innodb_adaptive_flushing=OFF simply flush innodb_io_capacity
per second whenever the dirty proportion of buffer pool pages exceeds
innodb_max_dirty_pages_pct_lwm. For innodb_adaptive_flushing=ON we try
to make page_cleaner_flush_pages_recommendation() more consistent and
predictable: if we are below innodb_adaptive_flushing_lwm, let us flush
pages according to the return value of af_get_pct_for_dirty().
innodb_max_dirty_pages_pct_lwm: Revert the change of the default value
that was made in MDEV-23399. The value innodb_max_dirty_pages_pct_lwm=0
guarantees that a shutdown of an idle server will be fast. Users might
be surprised if normal shutdown suddenly became slower when upgrading
within a GA release series.
innodb_checkpoint_usec: Remove. The master task will no longer perform
periodic log checkpoints. It is the duty of the page cleaner thread.
log_sys.max_modified_age: Remove. The current span of the
buf_pool.flush_list expressed in LSN only matters for adaptive
flushing (outside the 'furious flushing' condition).
For the correctness of checkpoints, the only thing that matters is
the checkpoint age (log_sys.lsn - log_sys.last_checkpoint_lsn).
This run-time constant was also reported as log_max_modified_age_sync.
log_sys.max_checkpoint_age_async: Remove. This does not serve any
purpose, because the checkpoints will now be triggered by the page
cleaner thread. We will retain the log_sys.max_checkpoint_age limit
for engaging 'furious flushing'.
page_cleaner.slot: Remove. It turns out that
page_cleaner_slot.flush_list_time was duplicating
page_cleaner.slot.flush_time and page_cleaner.slot.flush_list_pass
was duplicating page_cleaner.flush_pass.
Likewise, there were some redundant monitor counters, because the
page cleaner thread no longer performs any buf_pool.LRU flushing, and
because there only is one buf_flush_page_cleaner thread.
buf_flush_sync_lsn: Protect writes by buf_pool.flush_list_mutex.
buf_pool_t::get_oldest_modification(): Add a parameter to specify the
return value when no persistent data pages are dirty. Require the
caller to hold buf_pool.flush_list_mutex.
log_buf_pool_get_oldest_modification(): Take the fall-back LSN
as a parameter. All callers will also invoke log_sys.get_lsn().
log_preflush_pool_modified_pages(): Replaced with buf_flush_wait_flushed().
buf_flush_wait_flushed(): Implement two limits. If not enough buffer pool
has been flushed, signal the page cleaner (unless innodb_flush_sync=OFF)
and wait for the page cleaner to complete. If the page cleaner
thread is not running (which can be the case durign shutdown),
initiate the flush and wait for it directly.
buf_flush_ahead(): If innodb_flush_sync=ON (the default),
submit a new buf_flush_sync_lsn target for the page cleaner
but do not wait for the flushing to finish.
log_get_capacity(), log_get_max_modified_age_async(): Remove, to make
it easier to see that af_get_pct_for_lsn() is not acquiring any mutexes.
page_cleaner_flush_pages_recommendation(): Protect all access to
buf_pool.flush_list with buf_pool.flush_list_mutex. Previously there
were some race conditions in the calculation.
buf_flush_sync_for_checkpoint(): New function to process
buf_flush_sync_lsn in the page cleaner thread. At the end of
each batch, we try to wake up any blocked buf_flush_wait_flushed().
If everything up to buf_flush_sync_lsn has been flushed, we will
reset buf_flush_sync_lsn=0. The page cleaner thread will keep
'furious flushing' until the limit is reached. Any threads that
are waiting in buf_flush_wait_flushed() will be able to resume
as soon as their own limit has been satisfied.
buf_flush_page_cleaner: Prioritize buf_flush_sync_lsn and do not
sleep as long as it is set. Do not update any page_cleaner statistics
for this special mode of operation. In the normal mode
(buf_flush_sync_lsn is not set for innodb_flush_sync=ON),
try to wake up once per second. No longer check whether
srv_inc_activity_count() has been called. After each batch,
try to perform a log checkpoint, because the best chances for
the checkpoint LSN to advance by the maximum amount are upon
completing a flushing batch.
log_t: Move buf_free, max_buf_free possibly to the same cache line
with log_sys.mutex.
log_margin_checkpoint_age(): Simplify the logic, and replace
a 0.1-second sleep with a call to buf_flush_wait_flushed() to
initiate flushing. Moved to the same compilation unit
with the only caller.
log_close(): Clean up the calculations. (Should be no functional
change.) Return whether flush-ahead is needed. Moved to the same
compilation unit with the only caller.
mtr_t::finish_write(): Return whether flush-ahead is needed.
mtr_t::commit(): Invoke buf_flush_ahead() when needed. Let us avoid
external calls in mtr_t::commit() and make the logic easier to follow
by having related code in a single compilation unit. Also, we will
invoke srv_stats.log_write_requests.inc() only once per
mini-transaction commit, while not holding mutexes.
log_checkpoint_margin(): Only care about log_sys.max_checkpoint_age.
Upon reaching log_sys.max_checkpoint_age where we must wait to prevent
the log from getting corrupted, let us wait for at most 1MiB of LSN
at a time, before rechecking the condition. This should allow writers
to proceed even if the redo log capacity has been reached and
'furious flushing' is in progress. We no longer care about
log_sys.max_modified_age_sync or log_sys.max_modified_age_async.
The log_sys.max_modified_age_sync could be a relic from the time when
there was a srv_master_thread that wrote dirty pages to data files.
Also, we no longer have any log_sys.max_checkpoint_age_async limit,
because log checkpoints will now be triggered by the page cleaner
thread upon completing buf_flush_lists().
log_set_capacity(): Simplify the calculations of the limit
(no functional change).
log_checkpoint_low(): Split from log_checkpoint(). Moved to the
same compilation unit with the caller.
log_make_checkpoint(): Only wait for everything to be flushed until
the current LSN.
create_log_file(): After checkpoint, invoke log_write_up_to()
to ensure that the FILE_CHECKPOINT record has been written.
This avoids ut_ad(!srv_log_file_created) in create_log_file_rename().
srv_start(): Do not call recv_recovery_from_checkpoint_start()
if the log has just been created. Set fil_system.space_id_reuse_warned
before dict_boot() has been executed, and clear it after recovery
has finished.
dict_boot(): Initialize fil_system.max_assigned_id.
srv_check_activity(): Remove. The activity count is counting transaction
commits and therefore mostly interesting for the purge of history.
BtrBulk::insert(): Do not explicitly wake up the page cleaner,
but do invoke srv_inc_activity_count(), because that counter is
still being used in buf_load_throttle_if_needed() for some
heuristics. (It might be cleaner to execute buf_load() in the
page cleaner thread!)
Reviewed by: Vladislav Vaintroub
2020-10-26 15:35:47 +01:00
|
|
|
fil_system.space_id_reuse_warned = true;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
/** Load tablespace.
|
|
|
|
|
|
|
|
@param[in] dirname directory name of the tablespace to open
|
|
|
|
@param[in] filname file name of the tablespece to open
|
|
|
|
@param[in] is_remote true if tablespace file is .isl
|
2020-12-02 17:29:49 +01:00
|
|
|
@param[in] skip_node_page0 true if we don't need to read node page 0. Otherwise
|
|
|
|
node page0 will be read, and it's size and free pages limit
|
|
|
|
will be set from page 0, what is neccessary for checking and fixing corrupted
|
|
|
|
pages.
|
2021-05-17 17:12:33 +02:00
|
|
|
@param[in] defer_space_id use the space id to create space object
|
|
|
|
when there is deferred tablespace
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
*/
|
|
|
|
static void xb_load_single_table_tablespace(const char *dirname,
|
|
|
|
const char *filname,
|
2020-12-02 17:29:49 +01:00
|
|
|
bool is_remote,
|
2021-05-17 17:12:33 +02:00
|
|
|
bool skip_node_page0,
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t defer_space_id)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
Follow-up fix to MDEV-12988 backup fails if innodb_undo_tablespaces>0
The fix broke mariabackup --prepare --incremental.
The restore of an incremental backup starts up (parts of) InnoDB twice.
First, all data files are discovered for applying .delta files. Then,
after the .delta files have been applied, InnoDB will be restarted
more completely, so that the redo log records will be applied via the
buffer pool.
During the first startup, the buffer pool is not initialized, and thus
trx_rseg_get_n_undo_tablespaces() must not be invoked. The apply of
the .delta files will currently assume that the --innodb-undo-tablespaces
option correctly specifies the number of undo tablespace files, just
like --backup does.
The second InnoDB startup of --prepare for applying the redo log will
properly invoke trx_rseg_get_n_undo_tablespaces().
enum srv_operation_mode: Add SRV_OPERATION_RESTORE_DELTA for
distinguishing the apply of .delta files from SRV_OPERATION_RESTORE.
srv_undo_tablespaces_init(): In mariabackup --prepare --incremental,
in the initial SRV_OPERATION_RESTORE_DELTA phase, do not invoke
trx_rseg_get_n_undo_tablespaces() because the buffer pool or the
redo logs are not available. Instead, blindly rely on the parameter
--innodb-undo-tablespaces.
2017-08-18 07:52:41 +02:00
|
|
|
ut_ad(srv_operation == SRV_OPERATION_BACKUP
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
|| srv_operation == SRV_OPERATION_RESTORE_DELTA
|
2022-01-29 13:40:25 +01:00
|
|
|
|| srv_operation == SRV_OPERATION_RESTORE
|
|
|
|
|| srv_operation == SRV_OPERATION_BACKUP_NO_DEFER);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/* Ignore .isl files on XtraBackup recovery. All tablespaces must be
|
|
|
|
local. */
|
Follow-up fix to MDEV-12988 backup fails if innodb_undo_tablespaces>0
The fix broke mariabackup --prepare --incremental.
The restore of an incremental backup starts up (parts of) InnoDB twice.
First, all data files are discovered for applying .delta files. Then,
after the .delta files have been applied, InnoDB will be restarted
more completely, so that the redo log records will be applied via the
buffer pool.
During the first startup, the buffer pool is not initialized, and thus
trx_rseg_get_n_undo_tablespaces() must not be invoked. The apply of
the .delta files will currently assume that the --innodb-undo-tablespaces
option correctly specifies the number of undo tablespace files, just
like --backup does.
The second InnoDB startup of --prepare for applying the redo log will
properly invoke trx_rseg_get_n_undo_tablespaces().
enum srv_operation_mode: Add SRV_OPERATION_RESTORE_DELTA for
distinguishing the apply of .delta files from SRV_OPERATION_RESTORE.
srv_undo_tablespaces_init(): In mariabackup --prepare --incremental,
in the initial SRV_OPERATION_RESTORE_DELTA phase, do not invoke
trx_rseg_get_n_undo_tablespaces() because the buffer pool or the
redo logs are not available. Instead, blindly rely on the parameter
--innodb-undo-tablespaces.
2017-08-18 07:52:41 +02:00
|
|
|
if (is_remote && srv_operation == SRV_OPERATION_RESTORE_DELTA) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
if (check_if_skip_table(filname)) {
|
|
|
|
return;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/* The name ends in .ibd or .isl;
|
|
|
|
try opening the file */
|
|
|
|
char* name;
|
|
|
|
size_t dirlen = dirname == NULL ? 0 : strlen(dirname);
|
|
|
|
size_t namelen = strlen(filname);
|
|
|
|
ulint pathlen = dirname == NULL ? namelen + 1: dirlen + namelen + 2;
|
|
|
|
dberr_t err;
|
|
|
|
fil_space_t *space;
|
2021-05-17 17:12:33 +02:00
|
|
|
bool defer = false;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
name = static_cast<char*>(ut_malloc_nokey(pathlen));
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (dirname != NULL) {
|
2017-11-11 22:07:24 +01:00
|
|
|
snprintf(name, pathlen, "%s/%s", dirname, filname);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
name[pathlen - 5] = 0;
|
|
|
|
} else {
|
2017-11-11 22:07:24 +01:00
|
|
|
snprintf(name, pathlen, "%s", filname);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
name[pathlen - 5] = 0;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2021-05-20 13:58:25 +02:00
|
|
|
const fil_space_t::name_type n{name, pathlen - 5};
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
Datafile *file;
|
|
|
|
|
|
|
|
if (is_remote) {
|
|
|
|
RemoteDatafile* rf = new RemoteDatafile();
|
2021-05-20 13:58:25 +02:00
|
|
|
if (!rf->open_link_file(n)) {
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
die("Can't open datafile %s", name);
|
|
|
|
}
|
|
|
|
file = rf;
|
|
|
|
} else {
|
|
|
|
file = new Datafile();
|
2021-05-20 13:58:25 +02:00
|
|
|
file->make_filepath(".", n, IBD);
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (file->open_read_only(true) != DB_SUCCESS) {
|
2019-01-15 22:47:54 +01:00
|
|
|
die("Can't open datafile %s", name);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2019-01-30 21:50:51 +01:00
|
|
|
for (int i = 0; i < 10; i++) {
|
2021-05-17 17:12:33 +02:00
|
|
|
file->m_defer = false;
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
err = file->validate_first_page();
|
2021-05-17 17:12:33 +02:00
|
|
|
|
|
|
|
if (file->m_defer) {
|
|
|
|
if (defer_space_id) {
|
|
|
|
defer = true;
|
|
|
|
file->set_space_id(defer_space_id);
|
|
|
|
file->set_flags(FSP_FLAGS_PAGE_SSIZE());
|
|
|
|
err = DB_SUCCESS;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
} else if (err != DB_CORRUPTION) {
|
2019-01-30 21:50:51 +01:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
my_sleep(1000);
|
|
|
|
}
|
|
|
|
|
2021-05-17 17:12:33 +02:00
|
|
|
if (!defer && file->m_defer) {
|
2022-01-29 13:40:25 +01:00
|
|
|
const char *file_path = file->filepath();
|
|
|
|
defer_space_names.insert(
|
|
|
|
filename_to_spacename(
|
|
|
|
file_path, strlen(file_path)));
|
2021-05-17 17:12:33 +02:00
|
|
|
delete file;
|
|
|
|
ut_free(name);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2019-01-30 21:50:51 +01:00
|
|
|
bool is_empty_file = file->exists() && file->is_empty_file();
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (err == DB_SUCCESS && file->space_id() != SRV_TMP_SPACE_ID) {
|
2020-10-26 14:59:30 +01:00
|
|
|
space = fil_space_t::create(
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
file->space_id(), file->flags(),
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
FIL_TYPE_TABLESPACE, NULL/* TODO: crypt_data */);
|
|
|
|
|
|
|
|
ut_a(space != NULL);
|
2021-05-17 17:12:33 +02:00
|
|
|
fil_node_t* node= space->add(
|
|
|
|
file->filepath(),
|
|
|
|
skip_node_page0 ? file->detach() : pfs_os_file_t(),
|
|
|
|
0, false, false);
|
|
|
|
node->deferred= defer;
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_lock(&fil_system.mutex);
|
2021-08-31 14:09:47 +02:00
|
|
|
if (!space->read_page0())
|
|
|
|
err= DB_CANNOT_OPEN_FILE;
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_unlock(&fil_system.mutex);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
Follow-up fix to MDEV-12988 backup fails if innodb_undo_tablespaces>0
The fix broke mariabackup --prepare --incremental.
The restore of an incremental backup starts up (parts of) InnoDB twice.
First, all data files are discovered for applying .delta files. Then,
after the .delta files have been applied, InnoDB will be restarted
more completely, so that the redo log records will be applied via the
buffer pool.
During the first startup, the buffer pool is not initialized, and thus
trx_rseg_get_n_undo_tablespaces() must not be invoked. The apply of
the .delta files will currently assume that the --innodb-undo-tablespaces
option correctly specifies the number of undo tablespace files, just
like --backup does.
The second InnoDB startup of --prepare for applying the redo log will
properly invoke trx_rseg_get_n_undo_tablespaces().
enum srv_operation_mode: Add SRV_OPERATION_RESTORE_DELTA for
distinguishing the apply of .delta files from SRV_OPERATION_RESTORE.
srv_undo_tablespaces_init(): In mariabackup --prepare --incremental,
in the initial SRV_OPERATION_RESTORE_DELTA phase, do not invoke
trx_rseg_get_n_undo_tablespaces() because the buffer pool or the
redo logs are not available. Instead, blindly rely on the parameter
--innodb-undo-tablespaces.
2017-08-18 07:52:41 +02:00
|
|
|
if (srv_operation == SRV_OPERATION_RESTORE_DELTA
|
|
|
|
|| xb_close_files) {
|
2018-03-28 08:29:14 +02:00
|
|
|
space->close();
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
delete file;
|
|
|
|
|
2019-01-30 21:50:51 +01:00
|
|
|
if (err != DB_SUCCESS && xtrabackup_backup && !is_empty_file) {
|
2021-08-26 15:13:48 +02:00
|
|
|
die("Failed to validate first page of the file %s, error %d",name, (int)err);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
2020-03-10 13:14:53 +01:00
|
|
|
|
|
|
|
ut_free(name);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
|
|
|
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
static void xb_load_single_table_tablespace(const std::string &space_name,
|
2021-05-17 17:12:33 +02:00
|
|
|
bool skip_node_page0,
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t defer_space_id)
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
{
|
|
|
|
std::string name(space_name);
|
|
|
|
bool is_remote= access((name + ".ibd").c_str(), R_OK) != 0;
|
|
|
|
const char *extension= is_remote ? ".isl" : ".ibd";
|
|
|
|
name.append(extension);
|
|
|
|
char buf[FN_REFLEN];
|
|
|
|
strncpy(buf, name.c_str(), sizeof buf - 1);
|
|
|
|
buf[sizeof buf - 1]= '\0';
|
|
|
|
const char *dbname= buf;
|
|
|
|
char *p= strchr(buf, '/');
|
2021-05-17 17:12:33 +02:00
|
|
|
if (!p)
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
die("Unexpected tablespace %s filename %s", space_name.c_str(),
|
|
|
|
name.c_str());
|
|
|
|
*p= 0;
|
|
|
|
const char *tablename= p + 1;
|
2020-12-02 17:29:49 +01:00
|
|
|
xb_load_single_table_tablespace(dbname, tablename, is_remote,
|
2021-05-17 17:12:33 +02:00
|
|
|
skip_node_page0, defer_space_id);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
}
|
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
#ifdef _WIN32
|
|
|
|
/**
|
|
|
|
The os_file_opendir() function opens a directory stream corresponding to the
|
|
|
|
directory named by the dirname argument. The directory stream is positioned
|
|
|
|
at the first entry. In both Unix and Windows we automatically skip the '.'
|
|
|
|
and '..' items at the start of the directory listing.
|
|
|
|
@param[in] dirname directory name; it must not contain a trailing
|
|
|
|
'\' or '/'
|
|
|
|
@return directory stream, NULL if error */
|
|
|
|
os_file_dir_t os_file_opendir(const char *dirname)
|
|
|
|
{
|
|
|
|
char path[OS_FILE_MAX_PATH + 3];
|
|
|
|
|
|
|
|
ut_a(strlen(dirname) < OS_FILE_MAX_PATH);
|
|
|
|
|
|
|
|
strcpy(path, dirname);
|
|
|
|
strcpy(path + strlen(path), "\\*");
|
|
|
|
|
|
|
|
/* Note that in Windows opening the 'directory stream' also retrieves
|
|
|
|
the first entry in the directory. Since it is '.', that is no problem,
|
|
|
|
as we will skip over the '.' and '..' entries anyway. */
|
|
|
|
|
|
|
|
LPWIN32_FIND_DATA lpFindFileData= static_cast<LPWIN32_FIND_DATA>
|
|
|
|
(ut_malloc_nokey(sizeof(WIN32_FIND_DATA)));
|
|
|
|
os_file_dir_t dir= FindFirstFile((LPCTSTR) path, lpFindFileData);
|
|
|
|
ut_free(lpFindFileData);
|
|
|
|
|
|
|
|
return dir;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/** This function returns information of the next file in the directory. We jump
|
|
|
|
over the '.' and '..' entries in the directory.
|
|
|
|
@param[in] dirname directory name or path
|
|
|
|
@param[in] dir directory stream
|
|
|
|
@param[out] info buffer where the info is returned
|
|
|
|
@return 0 if ok, -1 if error, 1 if at the end of the directory */
|
|
|
|
int
|
|
|
|
os_file_readdir_next_file(
|
|
|
|
const char* dirname,
|
|
|
|
os_file_dir_t dir,
|
|
|
|
os_file_stat_t* info)
|
|
|
|
{
|
|
|
|
#ifdef _WIN32
|
|
|
|
BOOL ret;
|
|
|
|
int status;
|
|
|
|
WIN32_FIND_DATA find_data;
|
|
|
|
|
|
|
|
next_file:
|
|
|
|
ret = FindNextFile(dir, &find_data);
|
|
|
|
|
|
|
|
if (ret > 0) {
|
|
|
|
|
|
|
|
const char* name;
|
|
|
|
|
|
|
|
name = static_cast<const char*>(find_data.cFileName);
|
|
|
|
|
|
|
|
ut_a(strlen(name) < OS_FILE_MAX_PATH);
|
|
|
|
|
|
|
|
if (strcmp(name, ".") == 0 || strcmp(name, "..") == 0) {
|
|
|
|
|
|
|
|
goto next_file;
|
|
|
|
}
|
|
|
|
|
|
|
|
strcpy(info->name, name);
|
|
|
|
|
|
|
|
info->size = find_data.nFileSizeHigh;
|
|
|
|
info->size <<= 32;
|
|
|
|
info->size |= find_data.nFileSizeLow;
|
|
|
|
|
|
|
|
if (find_data.dwFileAttributes
|
|
|
|
& FILE_ATTRIBUTE_REPARSE_POINT) {
|
|
|
|
|
|
|
|
/* TODO: test Windows symlinks */
|
|
|
|
/* TODO: MySQL has apparently its own symlink
|
|
|
|
implementation in Windows, dbname.sym can
|
|
|
|
redirect a database directory:
|
|
|
|
REFMAN "windows-symbolic-links.html" */
|
|
|
|
|
|
|
|
info->type = OS_FILE_TYPE_LINK;
|
|
|
|
|
|
|
|
} else if (find_data.dwFileAttributes
|
|
|
|
& FILE_ATTRIBUTE_DIRECTORY) {
|
|
|
|
|
|
|
|
info->type = OS_FILE_TYPE_DIR;
|
|
|
|
|
|
|
|
} else {
|
|
|
|
|
|
|
|
/* It is probably safest to assume that all other
|
|
|
|
file types are normal. Better to check them rather
|
|
|
|
than blindly skip them. */
|
|
|
|
|
|
|
|
info->type = OS_FILE_TYPE_FILE;
|
|
|
|
}
|
|
|
|
|
|
|
|
status = 0;
|
|
|
|
|
|
|
|
} else {
|
|
|
|
DWORD err = GetLastError();
|
|
|
|
if (err == ERROR_NO_MORE_FILES) {
|
|
|
|
status = 1;
|
|
|
|
} else {
|
2021-08-11 11:40:42 +02:00
|
|
|
msg("FindNextFile in %s returned %lu", dirname, err);
|
2021-05-18 11:13:18 +02:00
|
|
|
status = -1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return(status);
|
|
|
|
#else
|
|
|
|
struct dirent* ent;
|
|
|
|
char* full_path;
|
|
|
|
int ret;
|
|
|
|
struct stat statinfo;
|
|
|
|
|
|
|
|
next_file:
|
|
|
|
|
|
|
|
ent = readdir(dir);
|
|
|
|
|
|
|
|
if (ent == NULL) {
|
|
|
|
|
|
|
|
return(1);
|
|
|
|
}
|
|
|
|
|
|
|
|
ut_a(strlen(ent->d_name) < OS_FILE_MAX_PATH);
|
|
|
|
|
|
|
|
if (strcmp(ent->d_name, ".") == 0 || strcmp(ent->d_name, "..") == 0) {
|
|
|
|
|
|
|
|
goto next_file;
|
|
|
|
}
|
|
|
|
|
|
|
|
strcpy(info->name, ent->d_name);
|
|
|
|
|
|
|
|
full_path = static_cast<char*>(
|
|
|
|
ut_malloc_nokey(strlen(dirname) + strlen(ent->d_name) + 10));
|
2021-05-19 07:39:39 +02:00
|
|
|
if (!full_path) {
|
|
|
|
return -1;
|
|
|
|
}
|
2021-05-18 11:13:18 +02:00
|
|
|
|
|
|
|
sprintf(full_path, "%s/%s", dirname, ent->d_name);
|
|
|
|
|
|
|
|
ret = stat(full_path, &statinfo);
|
|
|
|
|
|
|
|
if (ret) {
|
|
|
|
|
|
|
|
if (errno == ENOENT) {
|
|
|
|
/* readdir() returned a file that does not exist,
|
|
|
|
it must have been deleted in the meantime. Do what
|
|
|
|
would have happened if the file was deleted before
|
|
|
|
readdir() - ignore and go to the next entry.
|
|
|
|
If this is the last entry then info->name will still
|
|
|
|
contain the name of the deleted file when this
|
|
|
|
function returns, but this is not an issue since the
|
|
|
|
caller shouldn't be looking at info when end of
|
|
|
|
directory is returned. */
|
|
|
|
|
|
|
|
ut_free(full_path);
|
|
|
|
|
|
|
|
goto next_file;
|
|
|
|
}
|
|
|
|
|
|
|
|
msg("stat %s: Got error %d", full_path, errno);
|
|
|
|
|
|
|
|
ut_free(full_path);
|
|
|
|
|
|
|
|
return(-1);
|
|
|
|
}
|
|
|
|
|
|
|
|
info->size = statinfo.st_size;
|
|
|
|
|
|
|
|
if (S_ISDIR(statinfo.st_mode)) {
|
|
|
|
info->type = OS_FILE_TYPE_DIR;
|
|
|
|
} else if (S_ISLNK(statinfo.st_mode)) {
|
|
|
|
info->type = OS_FILE_TYPE_LINK;
|
|
|
|
} else if (S_ISREG(statinfo.st_mode)) {
|
|
|
|
info->type = OS_FILE_TYPE_FILE;
|
|
|
|
} else {
|
|
|
|
info->type = OS_FILE_TYPE_UNKNOWN;
|
|
|
|
}
|
|
|
|
|
|
|
|
ut_free(full_path);
|
|
|
|
return(0);
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************//**
|
|
|
|
A fault-tolerant function that tries to read the next file name in the
|
|
|
|
directory. We retry 100 times if os_file_readdir_next_file() returns -1. The
|
|
|
|
idea is to read as much good data as we can and jump over bad data.
|
|
|
|
@return 0 if ok, -1 if error even after the retries, 1 if at the end
|
|
|
|
of the directory */
|
|
|
|
int
|
|
|
|
fil_file_readdir_next_file(
|
|
|
|
/*=======================*/
|
|
|
|
dberr_t* err, /*!< out: this is set to DB_ERROR if an error
|
|
|
|
was encountered, otherwise not changed */
|
|
|
|
const char* dirname,/*!< in: directory name or path */
|
|
|
|
os_file_dir_t dir, /*!< in: directory stream */
|
|
|
|
os_file_stat_t* info) /*!< in/out: buffer where the
|
|
|
|
info is returned */
|
|
|
|
{
|
|
|
|
for (ulint i = 0; i < 100; i++) {
|
|
|
|
int ret = os_file_readdir_next_file(dirname, dir, info);
|
|
|
|
|
|
|
|
if (ret != -1) {
|
|
|
|
|
|
|
|
return(ret);
|
|
|
|
}
|
|
|
|
|
|
|
|
ib::error() << "os_file_readdir_next_file() returned -1 in"
|
|
|
|
" directory " << dirname
|
|
|
|
<< ", crash recovery may have failed"
|
|
|
|
" for some .ibd files!";
|
|
|
|
|
|
|
|
*err = DB_ERROR;
|
|
|
|
}
|
|
|
|
|
|
|
|
return(-1);
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/** Scan the database directories under the MySQL datadir, looking for
|
|
|
|
.ibd files and determining the space id in each of them.
|
|
|
|
@return DB_SUCCESS or error number */
|
2017-09-08 00:53:21 +02:00
|
|
|
|
|
|
|
static dberr_t enumerate_ibd_files(process_single_tablespace_func_t callback)
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
char* dbpath = NULL;
|
|
|
|
ulint dbpath_len = 100;
|
|
|
|
os_file_dir_t dir;
|
|
|
|
os_file_dir_t dbdir;
|
|
|
|
os_file_stat_t dbinfo;
|
|
|
|
os_file_stat_t fileinfo;
|
|
|
|
dberr_t err = DB_SUCCESS;
|
2017-09-14 18:42:09 +02:00
|
|
|
size_t len;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
|
|
|
/* The datadir of MySQL is always the default directory of mysqld */
|
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
dir = os_file_opendir(fil_path_to_mysql_datadir);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
if (UNIV_UNLIKELY(dir == IF_WIN(INVALID_HANDLE_VALUE, nullptr))) {
|
|
|
|
msg("cannot open dir %s", fil_path_to_mysql_datadir);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(DB_ERROR);
|
|
|
|
}
|
|
|
|
|
|
|
|
dbpath = static_cast<char*>(ut_malloc_nokey(dbpath_len));
|
|
|
|
|
|
|
|
/* Scan all directories under the datadir. They are the database
|
|
|
|
directories of MySQL. */
|
|
|
|
|
|
|
|
ret = fil_file_readdir_next_file(&err, fil_path_to_mysql_datadir, dir,
|
|
|
|
&dbinfo);
|
|
|
|
while (ret == 0) {
|
|
|
|
|
|
|
|
/* General tablespaces are always at the first level of the
|
|
|
|
data home dir */
|
2021-05-17 17:12:33 +02:00
|
|
|
if (dbinfo.type != OS_FILE_TYPE_FILE) {
|
|
|
|
const bool is_isl = ends_with(dbinfo.name, ".isl");
|
|
|
|
if (is_isl || ends_with(dbinfo.name,".ibd")) {
|
|
|
|
(*callback)(nullptr, dbinfo.name, is_isl,
|
|
|
|
false, 0);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (dbinfo.type == OS_FILE_TYPE_FILE
|
|
|
|
|| dbinfo.type == OS_FILE_TYPE_UNKNOWN) {
|
|
|
|
|
|
|
|
goto next_datadir_item;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* We found a symlink or a directory; try opening it to see
|
|
|
|
if a symlink is a directory */
|
|
|
|
|
|
|
|
len = strlen(fil_path_to_mysql_datadir)
|
|
|
|
+ strlen (dbinfo.name) + 2;
|
|
|
|
if (len > dbpath_len) {
|
|
|
|
dbpath_len = len;
|
|
|
|
|
|
|
|
if (dbpath) {
|
|
|
|
ut_free(dbpath);
|
|
|
|
}
|
|
|
|
|
|
|
|
dbpath = static_cast<char*>(ut_malloc_nokey(dbpath_len));
|
|
|
|
}
|
2017-11-11 22:07:24 +01:00
|
|
|
snprintf(dbpath, dbpath_len,
|
|
|
|
"%s/%s", fil_path_to_mysql_datadir, dbinfo.name);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
|
|
|
if (check_if_skip_database_by_path(dbpath)) {
|
|
|
|
fprintf(stderr, "Skipping db: %s\n", dbpath);
|
|
|
|
goto next_datadir_item;
|
|
|
|
}
|
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
dbdir = os_file_opendir(dbpath);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
if (UNIV_UNLIKELY(dbdir != IF_WIN(INVALID_HANDLE_VALUE,NULL))){
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/* We found a database directory; loop through it,
|
|
|
|
looking for possible .ibd files in it */
|
|
|
|
|
|
|
|
for (ret = fil_file_readdir_next_file(&err, dbpath,
|
|
|
|
dbdir,
|
|
|
|
&fileinfo);
|
|
|
|
ret == 0;
|
|
|
|
ret = fil_file_readdir_next_file(&err, dbpath,
|
|
|
|
dbdir,
|
|
|
|
&fileinfo)) {
|
|
|
|
if (fileinfo.type == OS_FILE_TYPE_DIR) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* We found a symlink or a file */
|
2017-09-14 18:42:09 +02:00
|
|
|
if (strlen(fileinfo.name) > 4) {
|
2017-09-13 14:10:48 +02:00
|
|
|
bool is_isl= false;
|
2017-11-08 12:12:11 +01:00
|
|
|
if (ends_with(fileinfo.name, ".ibd") || ((is_isl = ends_with(fileinfo.name, ".isl"))))
|
2021-05-17 17:12:33 +02:00
|
|
|
(*callback)(dbinfo.name, fileinfo.name, is_isl, false, 0);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
if (os_file_closedir_failed(dbdir)) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
fprintf(stderr, "InnoDB: Warning: could not"
|
2017-09-13 14:10:48 +02:00
|
|
|
" close database directory %s\n",
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
dbpath);
|
|
|
|
|
|
|
|
err = DB_ERROR;
|
|
|
|
}
|
|
|
|
|
|
|
|
} else {
|
2021-05-18 11:13:18 +02:00
|
|
|
msg("Can't open dir %s", dbpath);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
err = DB_ERROR;
|
|
|
|
break;
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
next_datadir_item:
|
|
|
|
ret = fil_file_readdir_next_file(&err,
|
|
|
|
fil_path_to_mysql_datadir,
|
|
|
|
dir, &dbinfo);
|
|
|
|
}
|
|
|
|
|
|
|
|
ut_free(dbpath);
|
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
if (os_file_closedir_failed(dir)) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
fprintf(stderr,
|
2021-03-18 08:47:30 +01:00
|
|
|
"InnoDB: Error: could not close MariaDB datadir\n");
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(DB_ERROR);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(err);
|
|
|
|
}
|
|
|
|
|
2022-12-02 07:24:35 +01:00
|
|
|
/** Close all undo tablespaces while applying incremental delta */
|
|
|
|
static void xb_close_undo_tablespaces()
|
|
|
|
{
|
|
|
|
if (srv_undo_space_id_start == 0)
|
|
|
|
return;
|
2022-12-07 08:39:46 +01:00
|
|
|
for (uint32_t space_id= srv_undo_space_id_start;
|
2022-12-02 07:24:35 +01:00
|
|
|
space_id < srv_undo_space_id_start + srv_undo_tablespaces_open;
|
|
|
|
space_id++)
|
|
|
|
{
|
|
|
|
fil_space_t *space= fil_space_get(space_id);
|
|
|
|
ut_ad(space);
|
|
|
|
space->close();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/****************************************************************************
|
|
|
|
Populates the tablespace memory cache by scanning for and opening data files.
|
|
|
|
@returns DB_SUCCESS or error code.*/
|
|
|
|
static
|
|
|
|
dberr_t
|
|
|
|
xb_load_tablespaces()
|
|
|
|
{
|
|
|
|
bool create_new_db;
|
|
|
|
dberr_t err;
|
|
|
|
ulint sum_of_new_sizes;
|
|
|
|
|
|
|
|
ut_ad(srv_operation == SRV_OPERATION_BACKUP
|
Follow-up fix to MDEV-12988 backup fails if innodb_undo_tablespaces>0
The fix broke mariabackup --prepare --incremental.
The restore of an incremental backup starts up (parts of) InnoDB twice.
First, all data files are discovered for applying .delta files. Then,
after the .delta files have been applied, InnoDB will be restarted
more completely, so that the redo log records will be applied via the
buffer pool.
During the first startup, the buffer pool is not initialized, and thus
trx_rseg_get_n_undo_tablespaces() must not be invoked. The apply of
the .delta files will currently assume that the --innodb-undo-tablespaces
option correctly specifies the number of undo tablespace files, just
like --backup does.
The second InnoDB startup of --prepare for applying the redo log will
properly invoke trx_rseg_get_n_undo_tablespaces().
enum srv_operation_mode: Add SRV_OPERATION_RESTORE_DELTA for
distinguishing the apply of .delta files from SRV_OPERATION_RESTORE.
srv_undo_tablespaces_init(): In mariabackup --prepare --incremental,
in the initial SRV_OPERATION_RESTORE_DELTA phase, do not invoke
trx_rseg_get_n_undo_tablespaces() because the buffer pool or the
redo logs are not available. Instead, blindly rely on the parameter
--innodb-undo-tablespaces.
2017-08-18 07:52:41 +02:00
|
|
|
|| srv_operation == SRV_OPERATION_RESTORE_DELTA);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
|
|
|
err = srv_sys_space.check_file_spec(&create_new_db, 0);
|
|
|
|
|
|
|
|
/* create_new_db must not be true. */
|
|
|
|
if (err != DB_SUCCESS || create_new_db) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Could not find data files at the specified datadir");
|
2017-04-18 20:43:20 +02:00
|
|
|
return(DB_ERROR);
|
|
|
|
}
|
|
|
|
|
2018-12-18 17:31:52 +01:00
|
|
|
for (int i= 0; i < 10; i++) {
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
err = srv_sys_space.open_or_create(false, false, &sum_of_new_sizes);
|
2018-12-18 17:31:52 +01:00
|
|
|
if (err == DB_PAGE_CORRUPTED || err == DB_CORRUPTION) {
|
|
|
|
my_sleep(1000);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
break;
|
|
|
|
}
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
|
|
|
if (err != DB_SUCCESS) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Could not open data files.\n");
|
2017-04-18 20:43:20 +02:00
|
|
|
return(err);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Add separate undo tablespaces to fil_system */
|
2022-10-24 17:16:43 +02:00
|
|
|
err = srv_undo_tablespaces_init(false, nullptr);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
if (err != DB_SUCCESS) {
|
|
|
|
return(err);
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/* It is important to call xb_load_single_table_tablespaces() after
|
2017-04-18 20:43:20 +02:00
|
|
|
srv_undo_tablespaces_init(), because fil_is_user_tablespace_id() *
|
|
|
|
relies on srv_undo_tablespaces_open to be properly initialized */
|
|
|
|
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: Generating a list of tablespaces");
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-09-08 00:53:21 +02:00
|
|
|
err = enumerate_ibd_files(xb_load_single_table_tablespace);
|
2017-04-18 20:43:20 +02:00
|
|
|
if (err != DB_SUCCESS) {
|
|
|
|
return(err);
|
|
|
|
}
|
|
|
|
|
2022-12-02 07:24:35 +01:00
|
|
|
if (srv_operation == SRV_OPERATION_RESTORE_DELTA) {
|
|
|
|
xb_close_undo_tablespaces();
|
|
|
|
}
|
|
|
|
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
DBUG_MARIABACKUP_EVENT("after_load_tablespaces", {});
|
2017-04-18 20:43:20 +02:00
|
|
|
return(DB_SUCCESS);
|
|
|
|
}
|
|
|
|
|
2020-10-26 14:04:24 +01:00
|
|
|
/** Destroy the tablespace memory cache. */
|
|
|
|
static void xb_data_files_close()
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2020-10-26 15:04:12 +01:00
|
|
|
fil_space_t::close_all();
|
2020-10-26 14:04:24 +01:00
|
|
|
buf_dblwr.close();
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Allocate and initialize the entry for databases and tables filtering
|
|
|
|
hash tables. If memory allocation is not successful, terminate program.
|
|
|
|
@return pointer to the created entry. */
|
|
|
|
static
|
|
|
|
xb_filter_entry_t *
|
|
|
|
xb_new_filter_entry(
|
|
|
|
/*================*/
|
|
|
|
const char* name) /*!< in: name of table/database */
|
|
|
|
{
|
|
|
|
xb_filter_entry_t *entry;
|
|
|
|
ulint namelen = strlen(name);
|
|
|
|
|
|
|
|
ut_a(namelen <= NAME_LEN * 2 + 1);
|
|
|
|
|
|
|
|
entry = static_cast<xb_filter_entry_t *>
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
(malloc(sizeof(xb_filter_entry_t) + namelen + 1));
|
2017-04-18 20:43:20 +02:00
|
|
|
memset(entry, '\0', sizeof(xb_filter_entry_t) + namelen + 1);
|
|
|
|
entry->name = ((char*)entry) + sizeof(xb_filter_entry_t);
|
|
|
|
strcpy(entry->name, name);
|
|
|
|
entry->has_tables = FALSE;
|
|
|
|
|
|
|
|
return entry;
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Add entry to hash table. If hash table is NULL, allocate and initialize
|
|
|
|
new hash table */
|
|
|
|
static
|
|
|
|
xb_filter_entry_t*
|
|
|
|
xb_add_filter(
|
|
|
|
const char* name, /*!< in: name of table/database */
|
2020-06-18 11:26:28 +02:00
|
|
|
hash_table_t* hash) /*!< in/out: hash to insert into */
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2021-06-14 08:15:20 +02:00
|
|
|
xb_filter_entry_t* entry = xb_new_filter_entry(name);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
if (UNIV_UNLIKELY(!hash->array)) {
|
|
|
|
hash->create(1000);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2021-06-14 08:15:20 +02:00
|
|
|
const ulint fold = my_crc32c(0, entry->name, strlen(entry->name));
|
|
|
|
HASH_INSERT(xb_filter_entry_t, name_hash, hash, fold, entry);
|
2017-04-18 20:43:20 +02:00
|
|
|
return entry;
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Validate name of table or database. If name is invalid, program will
|
|
|
|
be finished with error code */
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_validate_name(
|
|
|
|
/*=============*/
|
|
|
|
const char* name, /*!< in: name */
|
|
|
|
size_t len) /*!< in: length of name */
|
|
|
|
{
|
|
|
|
const char* p;
|
|
|
|
|
|
|
|
/* perform only basic validation. validate length and
|
|
|
|
path symbols */
|
|
|
|
if (len > NAME_LEN) {
|
2019-01-15 22:47:54 +01:00
|
|
|
die("name `%s` is too long.", name);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
p = strpbrk(name, "/\\~");
|
2017-12-02 20:50:08 +01:00
|
|
|
if (p && (uint) (p - name) < NAME_LEN) {
|
2019-01-15 22:47:54 +01:00
|
|
|
die("name `%s` is not valid.", name);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Register new filter entry which can be either database
|
|
|
|
or table name. */
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_register_filter_entry(
|
|
|
|
/*=====================*/
|
2017-04-19 15:09:03 +02:00
|
|
|
const char* name, /*!< in: name */
|
2020-06-18 11:26:28 +02:00
|
|
|
hash_table_t* databases_hash,
|
|
|
|
hash_table_t* tables_hash
|
2017-04-19 15:09:03 +02:00
|
|
|
)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
|
|
|
const char* p;
|
|
|
|
size_t namelen;
|
|
|
|
xb_filter_entry_t* db_entry = NULL;
|
|
|
|
|
|
|
|
namelen = strlen(name);
|
|
|
|
if ((p = strchr(name, '.')) != NULL) {
|
|
|
|
char dbname[NAME_LEN + 1];
|
|
|
|
|
|
|
|
xb_validate_name(name, p - name);
|
|
|
|
xb_validate_name(p + 1, namelen - (p - name));
|
|
|
|
|
|
|
|
strncpy(dbname, name, p - name);
|
|
|
|
dbname[p - name] = 0;
|
|
|
|
|
2021-05-29 06:19:46 +02:00
|
|
|
if (databases_hash && databases_hash->array) {
|
2021-06-14 08:15:20 +02:00
|
|
|
const ulint fold = my_crc32c(0, dbname, p - name);
|
2020-06-18 11:26:28 +02:00
|
|
|
HASH_SEARCH(name_hash, databases_hash,
|
2021-06-14 08:15:20 +02:00
|
|
|
fold,
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_filter_entry_t*,
|
|
|
|
db_entry, (void) 0,
|
|
|
|
!strcmp(db_entry->name, dbname));
|
|
|
|
}
|
|
|
|
if (!db_entry) {
|
2017-04-19 15:09:03 +02:00
|
|
|
db_entry = xb_add_filter(dbname, databases_hash);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
db_entry->has_tables = TRUE;
|
2017-04-19 15:09:03 +02:00
|
|
|
xb_add_filter(name, tables_hash);
|
2017-04-18 20:43:20 +02:00
|
|
|
} else {
|
|
|
|
xb_validate_name(name, namelen);
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
xb_add_filter(name, databases_hash);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_register_include_filter_entry(
|
|
|
|
const char* name
|
|
|
|
)
|
|
|
|
{
|
|
|
|
xb_register_filter_entry(name, &databases_include_hash,
|
|
|
|
&tables_include_hash);
|
|
|
|
}
|
|
|
|
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_register_exclude_filter_entry(
|
|
|
|
const char* name
|
|
|
|
)
|
|
|
|
{
|
|
|
|
xb_register_filter_entry(name, &databases_exclude_hash,
|
|
|
|
&tables_exclude_hash);
|
|
|
|
}
|
|
|
|
|
2020-04-20 16:23:43 +02:00
|
|
|
void register_ignore_db_dirs_filter(const char *name)
|
|
|
|
{
|
|
|
|
xb_add_filter(name, &databases_exclude_hash);
|
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/***********************************************************************
|
|
|
|
Register new table for the filter. */
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_register_table(
|
|
|
|
/*==============*/
|
|
|
|
const char* name) /*!< in: name of table */
|
|
|
|
{
|
|
|
|
if (strchr(name, '.') == NULL) {
|
2019-01-15 22:47:54 +01:00
|
|
|
die("`%s` is not fully qualified name.", name);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
xb_register_include_filter_entry(name);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
static
|
|
|
|
void
|
2017-04-19 15:09:03 +02:00
|
|
|
xb_add_regex_to_list(
|
|
|
|
const char* regex, /*!< in: regex */
|
|
|
|
const char* error_context, /*!< in: context to error message */
|
|
|
|
regex_list_t* list) /*! in: list to put new regex to */
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
|
|
|
char errbuf[100];
|
|
|
|
int ret;
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
regex_t compiled_regex;
|
|
|
|
ret = regcomp(&compiled_regex, regex, REG_EXTENDED);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (ret != 0) {
|
2017-04-19 15:09:03 +02:00
|
|
|
regerror(ret, &compiled_regex, errbuf, sizeof(errbuf));
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: error: %s regcomp(%s): %s",
|
2017-04-19 15:09:03 +02:00
|
|
|
error_context, regex, errbuf);
|
2017-04-18 20:43:20 +02:00
|
|
|
exit(EXIT_FAILURE);
|
|
|
|
}
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
list->push_back(compiled_regex);
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Register new regex for the include filter. */
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_register_include_regex(
|
|
|
|
/*==============*/
|
|
|
|
const char* regex) /*!< in: regex */
|
|
|
|
{
|
|
|
|
xb_add_regex_to_list(regex, "tables", ®ex_include_list);
|
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Register new regex for the exclude filter. */
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_register_exclude_regex(
|
|
|
|
/*==============*/
|
|
|
|
const char* regex) /*!< in: regex */
|
|
|
|
{
|
|
|
|
xb_add_regex_to_list(regex, "tables-exclude", ®ex_exclude_list);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
typedef void (*insert_entry_func_t)(const char*);
|
|
|
|
|
2020-04-20 16:23:43 +02:00
|
|
|
/* Scan string and load filter entries from it.
|
|
|
|
@param[in] list string representing a list
|
|
|
|
@param[in] delimiters delimiters of entries
|
|
|
|
@param[in] ins callback to add entry */
|
|
|
|
void xb_load_list_string(char *list, const char *delimiters,
|
|
|
|
insert_entry_func_t ins)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2020-04-20 16:23:43 +02:00
|
|
|
char *p;
|
|
|
|
char *saveptr;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-04-20 16:23:43 +02:00
|
|
|
p= strtok_r(list, delimiters, &saveptr);
|
|
|
|
while (p)
|
|
|
|
{
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-04-20 16:23:43 +02:00
|
|
|
ins(p);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-04-20 16:23:43 +02:00
|
|
|
p= strtok_r(NULL, delimiters, &saveptr);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Scan file and load filter entries from it. */
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_load_list_file(
|
|
|
|
/*==============*/
|
|
|
|
const char* filename, /*!< in: name of file */
|
|
|
|
insert_entry_func_t ins) /*!< in: callback to add entry */
|
|
|
|
{
|
|
|
|
char name_buf[NAME_LEN*2+2];
|
|
|
|
FILE* fp;
|
|
|
|
|
|
|
|
/* read and store the filenames */
|
|
|
|
fp = fopen(filename, "r");
|
|
|
|
if (!fp) {
|
2019-01-15 22:47:54 +01:00
|
|
|
die("Can't open %s",
|
2017-04-18 20:43:20 +02:00
|
|
|
filename);
|
|
|
|
}
|
|
|
|
while (fgets(name_buf, sizeof(name_buf), fp) != NULL) {
|
|
|
|
char* p = strchr(name_buf, '\n');
|
|
|
|
if (p) {
|
|
|
|
*p = '\0';
|
|
|
|
} else {
|
2019-01-15 22:47:54 +01:00
|
|
|
die("`%s...` name is too long", name_buf);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
ins(name_buf);
|
|
|
|
}
|
|
|
|
|
|
|
|
fclose(fp);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_filters_init()
|
|
|
|
{
|
|
|
|
if (xtrabackup_databases) {
|
|
|
|
xb_load_list_string(xtrabackup_databases, " \t",
|
2017-04-19 15:09:03 +02:00
|
|
|
xb_register_include_filter_entry);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_databases_file) {
|
|
|
|
xb_load_list_file(xtrabackup_databases_file,
|
2017-04-19 15:09:03 +02:00
|
|
|
xb_register_include_filter_entry);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_databases_exclude) {
|
|
|
|
xb_load_list_string(xtrabackup_databases_exclude, " \t",
|
|
|
|
xb_register_exclude_filter_entry);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_tables) {
|
|
|
|
xb_load_list_string(xtrabackup_tables, ",",
|
2017-04-19 15:09:03 +02:00
|
|
|
xb_register_include_regex);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_tables_file) {
|
|
|
|
xb_load_list_file(xtrabackup_tables_file, xb_register_table);
|
|
|
|
}
|
2017-04-19 15:09:03 +02:00
|
|
|
|
|
|
|
if (xtrabackup_tables_exclude) {
|
|
|
|
xb_load_list_string(xtrabackup_tables_exclude, ",",
|
|
|
|
xb_register_exclude_regex);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_filter_hash_free(hash_table_t* hash)
|
|
|
|
{
|
|
|
|
ulint i;
|
|
|
|
|
|
|
|
/* free the hash elements */
|
2020-06-18 11:26:28 +02:00
|
|
|
for (i = 0; i < hash->n_cells; i++) {
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_filter_entry_t* table;
|
|
|
|
|
|
|
|
table = static_cast<xb_filter_entry_t *>
|
|
|
|
(HASH_GET_FIRST(hash, i));
|
|
|
|
|
|
|
|
while (table) {
|
|
|
|
xb_filter_entry_t* prev_table = table;
|
|
|
|
|
|
|
|
table = static_cast<xb_filter_entry_t *>
|
|
|
|
(HASH_GET_NEXT(name_hash, prev_table));
|
2021-06-14 08:15:20 +02:00
|
|
|
const ulint fold = my_crc32c(0, prev_table->name,
|
|
|
|
strlen(prev_table->name));
|
2017-04-18 20:43:20 +02:00
|
|
|
HASH_DELETE(xb_filter_entry_t, name_hash, hash,
|
2021-06-14 08:15:20 +02:00
|
|
|
fold, prev_table);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
free(prev_table);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
hash->free();
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2017-04-19 15:09:03 +02:00
|
|
|
static void xb_regex_list_free(regex_list_t* list)
|
|
|
|
{
|
|
|
|
while (list->size() > 0) {
|
|
|
|
xb_regfree(&list->front());
|
|
|
|
list->pop_front();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/************************************************************************
|
|
|
|
Destroy table filters for partial backup. */
|
|
|
|
static
|
|
|
|
void
|
|
|
|
xb_filters_free()
|
|
|
|
{
|
2017-04-19 15:09:03 +02:00
|
|
|
xb_regex_list_free(®ex_include_list);
|
|
|
|
xb_regex_list_free(®ex_exclude_list);
|
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
if (tables_include_hash.array) {
|
|
|
|
xb_filter_hash_free(&tables_include_hash);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
if (tables_exclude_hash.array) {
|
|
|
|
xb_filter_hash_free(&tables_exclude_hash);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
if (databases_include_hash.array) {
|
|
|
|
xb_filter_hash_free(&databases_include_hash);
|
2017-04-19 15:09:03 +02:00
|
|
|
}
|
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
if (databases_exclude_hash.array) {
|
|
|
|
xb_filter_hash_free(&databases_exclude_hash);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
#ifdef RLIMIT_NOFILE
|
|
|
|
/**
|
2017-04-18 20:43:20 +02:00
|
|
|
Set the open files limit. Based on set_max_open_files().
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
@param max_file_limit requested open files limit
|
2017-04-18 20:43:20 +02:00
|
|
|
@return the resulting open files limit. May be less or more than the requested
|
|
|
|
value. */
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
static ulong xb_set_max_open_files(rlim_t max_file_limit)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
|
|
|
struct rlimit rlimit;
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
rlim_t old_cur;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (getrlimit(RLIMIT_NOFILE, &rlimit)) {
|
|
|
|
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
old_cur = rlimit.rlim_cur;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (rlimit.rlim_cur == RLIM_INFINITY) {
|
|
|
|
|
|
|
|
rlimit.rlim_cur = max_file_limit;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (rlimit.rlim_cur >= max_file_limit) {
|
|
|
|
|
|
|
|
max_file_limit = rlimit.rlim_cur;
|
|
|
|
goto end;
|
|
|
|
}
|
|
|
|
|
|
|
|
rlimit.rlim_cur = rlimit.rlim_max = max_file_limit;
|
|
|
|
|
|
|
|
if (setrlimit(RLIMIT_NOFILE, &rlimit)) {
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
/* Use original value */
|
|
|
|
max_file_limit = static_cast<ulong>(old_cur);
|
2017-04-18 20:43:20 +02:00
|
|
|
} else {
|
|
|
|
|
|
|
|
rlimit.rlim_cur = 0; /* Safety if next call fails */
|
|
|
|
|
|
|
|
(void) getrlimit(RLIMIT_NOFILE, &rlimit);
|
|
|
|
|
|
|
|
if (rlimit.rlim_cur) {
|
|
|
|
|
|
|
|
/* If call didn't fail */
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
max_file_limit = rlimit.rlim_cur;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
end:
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
return static_cast<ulong>(max_file_limit);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
#else
|
2020-03-25 15:09:14 +01:00
|
|
|
# define xb_set_max_open_files(x) 0UL
|
2017-04-18 20:43:20 +02:00
|
|
|
#endif
|
|
|
|
|
2022-12-21 12:41:10 +01:00
|
|
|
static void stop_backup_threads()
|
2017-07-06 18:20:34 +02:00
|
|
|
{
|
2022-12-21 12:41:10 +01:00
|
|
|
mysql_cond_broadcast(&log_copying_stop);
|
|
|
|
|
|
|
|
if (log_copying_running || have_io_watching_thread)
|
2020-12-04 15:18:04 +01:00
|
|
|
{
|
2023-01-10 13:42:50 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2020-12-04 15:18:04 +01:00
|
|
|
fputs("mariabackup: Stopping log copying thread", stderr);
|
|
|
|
fflush(stderr);
|
2023-01-10 13:42:50 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2022-12-21 12:41:10 +01:00
|
|
|
while (log_copying_running || have_io_watching_thread)
|
2020-12-04 15:18:04 +01:00
|
|
|
{
|
2022-12-21 12:41:10 +01:00
|
|
|
mysql_cond_broadcast(&log_copying_stop);
|
2023-01-10 13:42:50 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2020-12-04 15:18:04 +01:00
|
|
|
putc('.', stderr);
|
|
|
|
fflush(stderr);
|
2021-03-16 14:09:41 +01:00
|
|
|
std::this_thread::sleep_for(std::chrono::milliseconds(200));
|
2023-01-10 13:42:50 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2020-12-04 15:18:04 +01:00
|
|
|
}
|
|
|
|
putc('\n', stderr);
|
|
|
|
}
|
2017-07-06 18:20:34 +02:00
|
|
|
|
2022-12-21 12:41:10 +01:00
|
|
|
mysql_cond_destroy(&log_copying_stop);
|
2017-07-06 18:20:34 +02:00
|
|
|
}
|
|
|
|
|
2017-08-18 13:42:18 +02:00
|
|
|
/** Implement the core of --backup
|
|
|
|
@return whether the operation succeeded */
|
2018-06-07 13:29:35 +02:00
|
|
|
static bool xtrabackup_backup_low()
|
2017-08-18 13:42:18 +02:00
|
|
|
{
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2018-06-07 13:29:35 +02:00
|
|
|
ut_ad(!metadata_to_lsn);
|
|
|
|
|
2017-08-18 13:42:18 +02:00
|
|
|
/* read the latest checkpoint lsn */
|
|
|
|
{
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
const lsn_t lsn = recv_sys.lsn;
|
|
|
|
if (recv_sys.find_checkpoint() == DB_SUCCESS
|
|
|
|
&& log_sys.is_latest()) {
|
|
|
|
metadata_to_lsn = log_sys.next_checkpoint_lsn;
|
2017-10-25 16:17:21 +02:00
|
|
|
msg("mariabackup: The latest check point"
|
2019-01-14 22:28:23 +01:00
|
|
|
" (for incremental): '" LSN_PF "'",
|
2017-08-18 13:42:18 +02:00
|
|
|
metadata_to_lsn);
|
|
|
|
} else {
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
msg("Error: recv_sys.find_checkpoint() failed.");
|
2017-08-18 13:42:18 +02:00
|
|
|
}
|
2020-12-04 15:18:04 +01:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
recv_sys.lsn = lsn;
|
2022-12-21 12:41:10 +01:00
|
|
|
stop_backup_threads();
|
2017-08-18 13:42:18 +02:00
|
|
|
}
|
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (metadata_to_lsn && xtrabackup_copy_logfile()) {
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2018-06-15 12:31:43 +02:00
|
|
|
ds_close(dst_log_file);
|
2018-06-15 13:09:15 +02:00
|
|
|
dst_log_file = NULL;
|
2018-06-15 12:31:43 +02:00
|
|
|
return false;
|
|
|
|
}
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2018-06-15 12:31:43 +02:00
|
|
|
|
2018-06-07 13:29:35 +02:00
|
|
|
if (ds_close(dst_log_file) || !metadata_to_lsn) {
|
2017-08-18 13:42:18 +02:00
|
|
|
dst_log_file = NULL;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
dst_log_file = NULL;
|
|
|
|
|
2023-01-02 08:43:59 +01:00
|
|
|
std::vector<uint32_t> failed_ids;
|
|
|
|
std::set_difference(
|
|
|
|
fail_undo_ids.begin(), fail_undo_ids.end(),
|
|
|
|
undo_trunc_ids.begin(), undo_trunc_ids.end(),
|
|
|
|
std::inserter(failed_ids, failed_ids.begin()));
|
|
|
|
|
|
|
|
for (uint32_t id : failed_ids) {
|
|
|
|
msg("mariabackup: Failed to read undo log "
|
|
|
|
"tablespace space id %d and there is no undo "
|
|
|
|
"tablespace truncation redo record.",
|
|
|
|
id);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (failed_ids.size() > 0) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2017-08-18 13:42:18 +02:00
|
|
|
if(!xtrabackup_incremental) {
|
|
|
|
strcpy(metadata_type, "full-backuped");
|
|
|
|
metadata_from_lsn = 0;
|
|
|
|
} else {
|
|
|
|
strcpy(metadata_type, "incremental");
|
|
|
|
metadata_from_lsn = incremental_lsn;
|
|
|
|
}
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
metadata_last_lsn = recv_sys.lsn;
|
2017-08-18 13:42:18 +02:00
|
|
|
|
|
|
|
if (!xtrabackup_stream_metadata(ds_meta)) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: failed to stream metadata.");
|
2017-08-18 13:42:18 +02:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
if (xtrabackup_extra_lsndir) {
|
|
|
|
char filename[FN_REFLEN];
|
|
|
|
|
|
|
|
sprintf(filename, "%s/%s", xtrabackup_extra_lsndir,
|
|
|
|
XTRABACKUP_METADATA_FILENAME);
|
|
|
|
if (!xtrabackup_write_metadata(filename)) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: failed to write metadata "
|
|
|
|
"to '%s'.", filename);
|
2017-08-18 13:42:18 +02:00
|
|
|
return false;
|
|
|
|
}
|
2018-02-15 22:08:18 +01:00
|
|
|
sprintf(filename, "%s/%s", xtrabackup_extra_lsndir,
|
|
|
|
XTRABACKUP_INFO);
|
2019-05-27 13:52:27 +02:00
|
|
|
if (!write_xtrabackup_info(mysql_connection, filename, false, false)) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: failed to write info "
|
|
|
|
"to '%s'.", filename);
|
2018-02-15 22:08:18 +01:00
|
|
|
return false;
|
|
|
|
}
|
2017-08-18 13:42:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/** Implement --backup
|
|
|
|
@return whether the operation succeeded */
|
2019-04-08 14:08:04 +02:00
|
|
|
static bool xtrabackup_backup_func()
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
|
|
|
MY_STAT stat_info;
|
|
|
|
uint i;
|
|
|
|
uint count;
|
2017-05-26 13:04:19 +02:00
|
|
|
pthread_mutex_t count_mutex;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
CorruptedPages corrupted_pages;
|
2017-04-18 20:43:20 +02:00
|
|
|
data_thread_ctxt_t *data_threads;
|
2018-08-09 16:06:52 +02:00
|
|
|
pthread_cond_init(&scanned_lsn_cond, NULL);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
#ifdef USE_POSIX_FADVISE
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("uses posix_fadvise().");
|
2017-04-18 20:43:20 +02:00
|
|
|
#endif
|
|
|
|
|
|
|
|
/* cd to datadir */
|
|
|
|
|
|
|
|
if (my_setwd(mysql_real_data_home,MYF(MY_WME)))
|
|
|
|
{
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("my_setwd() failed , %s", mysql_real_data_home);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(false);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("cd to %s", mysql_real_data_home);
|
2021-10-22 10:14:57 +02:00
|
|
|
xb_plugin_backup_init(mysql_connection);
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
msg("open files limit requested %lu, set to %lu",
|
|
|
|
xb_open_files_limit,
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_set_max_open_files(xb_open_files_limit));
|
|
|
|
|
|
|
|
mysql_data_home= mysql_data_home_buff;
|
|
|
|
mysql_data_home[0]=FN_CURLIB; // all paths are relative from here
|
|
|
|
mysql_data_home[1]=0;
|
|
|
|
|
2017-05-19 07:53:58 +02:00
|
|
|
srv_n_purge_threads = 1;
|
2017-04-18 20:43:20 +02:00
|
|
|
srv_read_only_mode = TRUE;
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
srv_operation = SRV_OPERATION_BACKUP;
|
2018-08-09 16:06:52 +02:00
|
|
|
log_file_op = backup_file_op;
|
2023-01-02 08:43:59 +01:00
|
|
|
undo_space_trunc = backup_undo_trunc;
|
2022-01-29 13:40:25 +01:00
|
|
|
first_page_init = backup_first_page_op;
|
2018-06-07 13:29:35 +02:00
|
|
|
metadata_to_lsn = 0;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* initialize components */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if(innodb_init_param()) {
|
|
|
|
fail:
|
2020-12-04 15:18:04 +01:00
|
|
|
if (log_copying_running) {
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2020-12-04 15:18:04 +01:00
|
|
|
metadata_to_lsn = 1;
|
2022-12-21 12:41:10 +01:00
|
|
|
stop_backup_threads();
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2020-12-04 15:18:04 +01:00
|
|
|
}
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
log_file_op = NULL;
|
2023-01-02 08:43:59 +01:00
|
|
|
undo_space_trunc = NULL;
|
2022-01-29 13:40:25 +01:00
|
|
|
first_page_init = NULL;
|
2018-06-07 13:29:35 +02:00
|
|
|
if (dst_log_file) {
|
|
|
|
ds_close(dst_log_file);
|
|
|
|
dst_log_file = NULL;
|
|
|
|
}
|
2018-03-28 08:00:06 +02:00
|
|
|
if (fil_system.is_initialised()) {
|
2018-02-15 18:40:14 +01:00
|
|
|
innodb_shutdown();
|
|
|
|
}
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(false);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (srv_buf_pool_size >= 1000 * 1024 * 1024) {
|
|
|
|
/* Here we still have srv_pool_size counted
|
|
|
|
in kilobytes (in 4.0 this was in bytes)
|
|
|
|
srv_boot() converts the value to
|
|
|
|
pages; if buffer pool is less than 1000 MB,
|
|
|
|
assume fewer threads. */
|
|
|
|
srv_max_n_threads = 50000;
|
|
|
|
|
|
|
|
} else if (srv_buf_pool_size >= 8 * 1024 * 1024) {
|
|
|
|
|
|
|
|
srv_max_n_threads = 10000;
|
|
|
|
} else {
|
|
|
|
srv_max_n_threads = 1000; /* saves several MB of memory,
|
|
|
|
especially in 64-bit
|
|
|
|
computers */
|
|
|
|
}
|
2019-10-29 22:37:12 +01:00
|
|
|
srv_thread_pool_init();
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/* Reset the system variables in the recovery module. */
|
|
|
|
trx_pool_init();
|
2019-05-24 15:19:38 +02:00
|
|
|
recv_sys.create();
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
xb_filters_init();
|
|
|
|
|
|
|
|
xb_fil_io_init();
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2020-12-14 14:27:03 +01:00
|
|
|
if (os_aio_init()) {
|
|
|
|
msg("Error: cannot initialize AIO subsystem");
|
|
|
|
goto fail;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
log_sys.create();
|
2017-04-18 20:43:20 +02:00
|
|
|
/* get current checkpoint_lsn */
|
2022-06-16 10:15:21 +02:00
|
|
|
{
|
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2022-06-16 10:15:21 +02:00
|
|
|
dberr_t err = recv_sys.find_checkpoint();
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2022-06-16 10:15:21 +02:00
|
|
|
if (err != DB_SUCCESS) {
|
|
|
|
msg("Error: cannot read redo log header");
|
|
|
|
} else if (!log_sys.is_latest()) {
|
|
|
|
msg("Error: cannot process redo log before "
|
|
|
|
"MariaDB 10.8");
|
|
|
|
err = DB_ERROR;
|
|
|
|
} else {
|
|
|
|
recv_needed_recovery = true;
|
|
|
|
}
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2022-06-16 10:15:21 +02:00
|
|
|
if (err != DB_SUCCESS) {
|
|
|
|
goto fail;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
/* create extra LSN dir if it does not exist. */
|
|
|
|
if (xtrabackup_extra_lsndir
|
|
|
|
&&!my_stat(xtrabackup_extra_lsndir,&stat_info,MYF(0))
|
|
|
|
&& (my_mkdir(xtrabackup_extra_lsndir,0777,MYF(0)) < 0)) {
|
|
|
|
msg("Error: cannot mkdir %d: %s\n",
|
|
|
|
my_errno, xtrabackup_extra_lsndir);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
goto fail;
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
/* create target dir if not exist */
|
|
|
|
if (!xtrabackup_stream_str && !my_stat(xtrabackup_target_dir,&stat_info,MYF(0))
|
|
|
|
&& (my_mkdir(xtrabackup_target_dir,0777,MYF(0)) < 0)){
|
|
|
|
msg("Error: cannot mkdir %d: %s\n",
|
|
|
|
my_errno, xtrabackup_target_dir);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
goto fail;
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
xtrabackup_init_datasinks();
|
|
|
|
|
|
|
|
if (!select_history()) {
|
2022-05-06 09:49:35 +02:00
|
|
|
goto fail;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* open the log file */
|
|
|
|
memset(&stat_info, 0, sizeof(MY_STAT));
|
2020-01-11 20:05:28 +01:00
|
|
|
dst_log_file = ds_open(ds_redo, LOG_FILE_NAME, &stat_info);
|
2017-04-18 20:43:20 +02:00
|
|
|
if (dst_log_file == NULL) {
|
2020-01-11 20:05:28 +01:00
|
|
|
msg("Error: failed to open the target stream for '%s'.",
|
|
|
|
LOG_FILE_NAME);
|
2022-05-06 09:49:35 +02:00
|
|
|
goto fail;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* label it */
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
recv_sys.file_checkpoint = log_sys.next_checkpoint_lsn;
|
|
|
|
log_hdr_init();
|
2019-05-16 10:04:39 +02:00
|
|
|
/* Write log header*/
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (ds_write(dst_log_file, log_hdr_buf, 12288)) {
|
2019-05-16 10:04:39 +02:00
|
|
|
msg("error: write to logfile failed");
|
2022-05-06 09:49:35 +02:00
|
|
|
goto fail;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2018-06-07 13:29:35 +02:00
|
|
|
log_copying_running = true;
|
2022-12-21 12:41:10 +01:00
|
|
|
|
|
|
|
mysql_cond_init(0, &log_copying_stop, nullptr);
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* start io throttle */
|
2022-12-21 12:41:10 +01:00
|
|
|
if (xtrabackup_throttle) {
|
2017-04-18 20:43:20 +02:00
|
|
|
io_ticket = xtrabackup_throttle;
|
2020-12-04 15:18:04 +01:00
|
|
|
have_io_watching_thread = true;
|
|
|
|
mysql_cond_init(0, &wait_throttle, nullptr);
|
2022-12-21 12:41:10 +01:00
|
|
|
std::thread(io_watching_thread).detach();
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2018-03-13 10:49:30 +01:00
|
|
|
/* Populate fil_system with tablespaces to copy */
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (dberr_t err = xb_load_tablespaces()) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("merror: xb_load_tablespaces() failed with"
|
|
|
|
" error %s.", ut_strerr(err));
|
2018-06-07 13:29:35 +02:00
|
|
|
log_copying_running = false;
|
2022-05-06 09:49:35 +02:00
|
|
|
goto fail;
|
2018-03-13 10:49:30 +01:00
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* copy log file by current position */
|
2018-04-03 13:13:36 +02:00
|
|
|
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
recv_sys.lsn = log_sys.next_checkpoint_lsn;
|
2020-12-04 15:18:04 +01:00
|
|
|
|
|
|
|
const bool log_copy_failed = xtrabackup_copy_logfile();
|
|
|
|
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2020-12-04 15:18:04 +01:00
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (log_copy_failed) {
|
|
|
|
log_copying_running = false;
|
2022-05-18 13:26:48 +02:00
|
|
|
goto fail;
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
DBUG_MARIABACKUP_EVENT("before_innodb_log_copy_thread_started", {});
|
2019-10-02 20:11:59 +02:00
|
|
|
|
2021-03-17 08:03:06 +01:00
|
|
|
std::thread(log_copying_thread).detach();
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* FLUSH CHANGED_PAGE_BITMAPS call */
|
|
|
|
if (!flush_changed_page_bitmaps()) {
|
2022-05-06 09:49:35 +02:00
|
|
|
goto fail;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
ut_a(xtrabackup_parallel > 0);
|
|
|
|
|
|
|
|
if (xtrabackup_parallel > 1) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: Starting %u threads for parallel data "
|
2019-01-14 22:28:23 +01:00
|
|
|
"files transfer", xtrabackup_parallel);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2018-01-23 10:12:25 +01:00
|
|
|
if (opt_lock_ddl_per_table) {
|
|
|
|
mdl_lock_all();
|
2018-08-09 16:06:52 +02:00
|
|
|
|
|
|
|
DBUG_EXECUTE_IF("check_mdl_lock_works",
|
|
|
|
dbug_start_query_thread("ALTER TABLE test.t ADD COLUMN mdl_lock_column int",
|
2018-12-10 14:01:46 +01:00
|
|
|
"Waiting for table metadata lock", 0, 0););
|
2018-01-23 10:12:25 +01:00
|
|
|
}
|
|
|
|
|
2021-03-19 23:31:01 +01:00
|
|
|
datafiles_iter_t it;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* Create data copying threads */
|
|
|
|
data_threads = (data_thread_ctxt_t *)
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
malloc(sizeof(data_thread_ctxt_t) * xtrabackup_parallel);
|
2017-04-18 20:43:20 +02:00
|
|
|
count = xtrabackup_parallel;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
pthread_mutex_init(&count_mutex, NULL);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
for (i = 0; i < (uint) xtrabackup_parallel; i++) {
|
2021-03-19 23:31:01 +01:00
|
|
|
data_threads[i].it = ⁢
|
2017-04-18 20:43:20 +02:00
|
|
|
data_threads[i].num = i+1;
|
|
|
|
data_threads[i].count = &count;
|
2018-09-20 20:27:59 +02:00
|
|
|
data_threads[i].count_mutex = &count_mutex;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
data_threads[i].corrupted_pages = &corrupted_pages;
|
2021-03-17 08:03:06 +01:00
|
|
|
std::thread(data_copy_thread_func, data_threads + i).detach();
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Wait for threads to exit */
|
|
|
|
while (1) {
|
2021-03-16 14:09:41 +01:00
|
|
|
std::this_thread::sleep_for(std::chrono::seconds(1));
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
pthread_mutex_lock(&count_mutex);
|
|
|
|
bool stop = count == 0;
|
|
|
|
pthread_mutex_unlock(&count_mutex);
|
|
|
|
if (stop) {
|
2017-04-18 20:43:20 +02:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
pthread_mutex_destroy(&count_mutex);
|
|
|
|
free(data_threads);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
bool ok = backup_start(corrupted_pages);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2017-08-18 13:42:18 +02:00
|
|
|
if (ok) {
|
|
|
|
ok = xtrabackup_backup_low();
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-08-18 13:42:18 +02:00
|
|
|
backup_release();
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2018-03-30 00:13:01 +02:00
|
|
|
DBUG_EXECUTE_IF("check_mdl_lock_works",
|
2020-12-04 15:18:04 +01:00
|
|
|
pthread_join(dbug_alter_thread, nullptr););
|
2018-03-30 00:13:01 +02:00
|
|
|
|
2017-08-18 13:42:18 +02:00
|
|
|
if (ok) {
|
|
|
|
backup_finish();
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
if (opt_log_innodb_page_corruption)
|
|
|
|
ok = corrupted_pages.print_to_file(MB_CORRUPTED_PAGES_FILE);
|
|
|
|
|
2017-08-18 13:42:18 +02:00
|
|
|
if (!ok) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
goto fail;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
if (changed_page_bitmap) {
|
|
|
|
xb_page_bitmap_deinit(changed_page_bitmap);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
xtrabackup_destroy_datasinks();
|
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
msg("Redo log (from LSN " LSN_PF " to " LSN_PF ") was copied.",
|
|
|
|
log_sys.next_checkpoint_lsn, recv_sys.lsn);
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_filters_free();
|
|
|
|
|
|
|
|
xb_data_files_close();
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/* Make sure that the latest checkpoint was included */
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (metadata_to_lsn > recv_sys.lsn) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: failed to copy enough redo log ("
|
|
|
|
"LSN=" LSN_PF "; checkpoint LSN=" LSN_PF ").",
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
recv_sys.lsn, metadata_to_lsn);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
goto fail;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
|
|
|
innodb_shutdown();
|
2018-08-09 16:06:52 +02:00
|
|
|
log_file_op = NULL;
|
2023-01-02 08:43:59 +01:00
|
|
|
undo_space_trunc = NULL;
|
2022-01-29 13:40:25 +01:00
|
|
|
first_page_init = NULL;
|
2018-08-09 16:06:52 +02:00
|
|
|
pthread_cond_destroy(&scanned_lsn_cond);
|
2021-03-04 23:09:22 +01:00
|
|
|
if (!corrupted_pages.empty()) {
|
|
|
|
ut_ad(opt_log_innodb_page_corruption);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
msg("Error: corrupted innodb pages are found and logged to "
|
|
|
|
MB_CORRUPTED_PAGES_FILE " file");
|
|
|
|
}
|
2021-03-04 23:09:22 +01:00
|
|
|
return(true);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
|
|
|
|
/**
|
|
|
|
This function handles DDL changes at the end of backup, under protection of
|
|
|
|
FTWRL. This ensures consistent backup in presence of DDL.
|
|
|
|
|
|
|
|
- New tables, that were created during backup, are now copied into backup.
|
|
|
|
Also, tablespaces with optimized (no redo loggin DDL) are re-copied into
|
|
|
|
backup. This tablespaces will get the extension ".new" in the backup
|
|
|
|
|
|
|
|
- Tables that were renamed during backup, are marked as renamed
|
|
|
|
For these, file <old_name>.ren will be created.
|
|
|
|
The content of the file is the new tablespace name.
|
|
|
|
|
|
|
|
- Tables that were deleted during backup, are marked as deleted
|
|
|
|
For these , an empty file <name>.del will be created
|
|
|
|
|
|
|
|
It is the responsibility of the prepare phase to deal with .new, .ren, and .del
|
|
|
|
files.
|
|
|
|
*/
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
void backup_fix_ddl(CorruptedPages &corrupted_pages)
|
2018-08-09 16:06:52 +02:00
|
|
|
{
|
|
|
|
std::set<std::string> dropped_tables;
|
|
|
|
std::map<std::string, std::string> renamed_tables;
|
2021-05-17 17:12:33 +02:00
|
|
|
space_id_to_name_t new_tables;
|
2018-08-09 16:06:52 +02:00
|
|
|
|
2018-09-14 10:35:18 +02:00
|
|
|
/* Disable further DDL on backed up tables (only needed for --no-lock).*/
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
2018-09-14 10:35:18 +02:00
|
|
|
log_file_op = backup_file_op_fail;
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
2018-09-14 10:35:18 +02:00
|
|
|
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
DBUG_MARIABACKUP_EVENT("backup_fix_ddl", {});
|
2018-09-14 10:35:18 +02:00
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
for (space_id_to_name_t::iterator iter = ddl_tracker.tables_in_backup.begin();
|
|
|
|
iter != ddl_tracker.tables_in_backup.end();
|
|
|
|
iter++) {
|
|
|
|
|
|
|
|
const std::string name = iter->second;
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t id = iter->first;
|
2018-08-09 16:06:52 +02:00
|
|
|
|
|
|
|
if (ddl_tracker.drops.find(id) != ddl_tracker.drops.end()) {
|
|
|
|
dropped_tables.insert(name);
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
corrupted_pages.drop_space(id);
|
2018-08-09 16:06:52 +02:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (ddl_tracker.id_to_name.find(id) == ddl_tracker.id_to_name.end()) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* tablespace was affected by DDL. */
|
|
|
|
const std::string new_name = ddl_tracker.id_to_name[id];
|
|
|
|
if (new_name != name) {
|
2020-01-29 14:13:08 +01:00
|
|
|
renamed_tables[name] = new_name;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
if (opt_log_innodb_page_corruption)
|
2020-12-02 17:29:49 +01:00
|
|
|
corrupted_pages.rename_space(id, new_name);
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Find tables that were created during backup (and not removed).*/
|
|
|
|
for(space_id_to_name_t::iterator iter = ddl_tracker.id_to_name.begin();
|
|
|
|
iter != ddl_tracker.id_to_name.end();
|
|
|
|
iter++) {
|
|
|
|
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t id = iter->first;
|
2018-08-09 16:06:52 +02:00
|
|
|
std::string name = iter->second;
|
|
|
|
|
|
|
|
if (ddl_tracker.tables_in_backup.find(id) != ddl_tracker.tables_in_backup.end()) {
|
|
|
|
/* already processed above */
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2022-01-29 13:40:25 +01:00
|
|
|
if (ddl_tracker.drops.find(id) == ddl_tracker.drops.end()
|
|
|
|
&& ddl_tracker.deferred_tables.find(id)
|
|
|
|
== ddl_tracker.deferred_tables.end()) {
|
2018-08-09 16:06:52 +02:00
|
|
|
dropped_tables.erase(name);
|
2021-05-17 17:12:33 +02:00
|
|
|
new_tables[id] = name;
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
if (opt_log_innodb_page_corruption)
|
|
|
|
corrupted_pages.drop_space(id);
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Mark tablespaces for rename
|
|
|
|
for (std::map<std::string, std::string>::iterator iter = renamed_tables.begin();
|
|
|
|
iter != renamed_tables.end(); ++iter) {
|
|
|
|
const std::string old_name = iter->first;
|
|
|
|
std::string new_name = iter->second;
|
|
|
|
backup_file_printf((old_name + ".ren").c_str(), "%s", new_name.c_str());
|
|
|
|
}
|
|
|
|
|
|
|
|
// Mark tablespaces for drop
|
|
|
|
for (std::set<std::string>::iterator iter = dropped_tables.begin();
|
|
|
|
iter != dropped_tables.end();
|
|
|
|
iter++) {
|
|
|
|
const std::string name(*iter);
|
|
|
|
backup_file_printf((name + ".del").c_str(), "%s", "");
|
|
|
|
}
|
|
|
|
|
|
|
|
// Load and copy new tables.
|
|
|
|
// Close all datanodes first, reload only new tables.
|
|
|
|
std::vector<fil_node_t *> all_nodes;
|
2021-03-19 23:31:01 +01:00
|
|
|
datafiles_iter_t it;
|
|
|
|
while (fil_node_t *node = datafiles_iter_next(&it)) {
|
2018-08-09 16:06:52 +02:00
|
|
|
all_nodes.push_back(node);
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < all_nodes.size(); i++) {
|
|
|
|
fil_node_t *n = all_nodes[i];
|
|
|
|
if (n->space->id == 0)
|
|
|
|
continue;
|
2018-08-16 07:54:58 +02:00
|
|
|
if (n->is_open()) {
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_lock(&fil_system.mutex);
|
2018-08-16 07:54:58 +02:00
|
|
|
n->close();
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_unlock(&fil_system.mutex);
|
2018-08-16 07:54:58 +02:00
|
|
|
}
|
2018-08-09 16:06:52 +02:00
|
|
|
fil_space_free(n->space->id, false);
|
|
|
|
}
|
|
|
|
|
2018-12-10 14:01:46 +01:00
|
|
|
DBUG_EXECUTE_IF("check_mdl_lock_works", DBUG_ASSERT(new_tables.size() == 0););
|
2021-05-17 17:12:33 +02:00
|
|
|
|
2022-01-29 13:40:25 +01:00
|
|
|
srv_operation = SRV_OPERATION_BACKUP_NO_DEFER;
|
|
|
|
|
|
|
|
/* Mariabackup detected the FILE_MODIFY or FILE_RENAME
|
|
|
|
for the deferred tablespace. So it needs to read the
|
|
|
|
tablespace again if innodb doesn't have page0 initialization
|
|
|
|
redo log for it */
|
|
|
|
for (space_id_to_name_t::iterator iter =
|
|
|
|
ddl_tracker.deferred_tables.begin();
|
|
|
|
iter != ddl_tracker.deferred_tables.end();
|
|
|
|
iter++) {
|
|
|
|
if (check_if_skip_table(iter->second.c_str())) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (first_page_init_ids.find(iter->first)
|
|
|
|
!= first_page_init_ids.end()) {
|
|
|
|
new_tables[iter->first] = iter->second.c_str();
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
xb_load_single_table_tablespace(iter->second, false);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Mariabackup doesn't detect any FILE_OP for the deferred
|
|
|
|
tablespace. There is a possiblity that page0 could've
|
|
|
|
been corrupted persistently in the disk */
|
|
|
|
for (auto space_name: defer_space_names) {
|
|
|
|
if (!check_if_skip_table(space_name.c_str())) {
|
|
|
|
xb_load_single_table_tablespace(
|
|
|
|
space_name, false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
srv_operation = SRV_OPERATION_BACKUP;
|
|
|
|
|
2021-05-17 17:12:33 +02:00
|
|
|
for (const auto &t : new_tables) {
|
|
|
|
if (!check_if_skip_table(t.second.c_str())) {
|
|
|
|
xb_load_single_table_tablespace(t.second, false,
|
|
|
|
t.first);
|
|
|
|
}
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
|
2021-03-19 23:31:01 +01:00
|
|
|
datafiles_iter_t it2;
|
2018-08-09 16:06:52 +02:00
|
|
|
|
2021-03-19 23:31:01 +01:00
|
|
|
while (fil_node_t *node = datafiles_iter_next(&it2)) {
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
if (!fil_is_user_tablespace_id(node->space->id))
|
2018-08-09 16:06:52 +02:00
|
|
|
continue;
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
std::string dest_name= filename_to_spacename(
|
|
|
|
node->name, strlen(node->name));
|
2018-08-09 16:06:52 +02:00
|
|
|
dest_name.append(".new");
|
2021-05-17 17:12:33 +02:00
|
|
|
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
xtrabackup_copy_datafile(node, 0, dest_name.c_str(), wf_write_through,
|
|
|
|
corrupted_pages);
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* ================= prepare ================= */
|
|
|
|
|
|
|
|
/***********************************************************************
|
|
|
|
Generates path to the meta file path from a given path to an incremental .delta
|
|
|
|
by replacing trailing ".delta" with ".meta", or returns error if 'delta_path'
|
|
|
|
does not end with the ".delta" character sequence.
|
|
|
|
@return TRUE on success, FALSE on error. */
|
|
|
|
static
|
|
|
|
ibool
|
|
|
|
get_meta_path(
|
|
|
|
const char *delta_path, /* in: path to a .delta file */
|
|
|
|
char *meta_path) /* out: path to the corresponding .meta
|
|
|
|
file */
|
|
|
|
{
|
|
|
|
size_t len = strlen(delta_path);
|
|
|
|
|
|
|
|
if (len <= 6 || strcmp(delta_path + len - 6, ".delta")) {
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
memcpy(meta_path, delta_path, len - 6);
|
|
|
|
strcpy(meta_path + len - 6, XB_DELTA_INFO_SUFFIX);
|
|
|
|
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
/****************************************************************//**
|
|
|
|
Create a new tablespace on disk and return the handle to its opened
|
|
|
|
file. Code adopted from fil_create_new_single_table_tablespace with
|
|
|
|
the main difference that only disk file is created without updating
|
|
|
|
the InnoDB in-memory dictionary data structures.
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
@return true on success, false on error. */
|
2017-04-18 20:43:20 +02:00
|
|
|
static
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
bool
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_space_create_file(
|
|
|
|
/*==================*/
|
|
|
|
const char* path, /*!<in: path to tablespace */
|
2021-07-22 10:22:47 +02:00
|
|
|
uint32_t space_id, /*!<in: space id */
|
|
|
|
uint32_t flags, /*!<in: tablespace flags */
|
2017-05-19 12:59:43 +02:00
|
|
|
pfs_os_file_t* file) /*!<out: file handle */
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
bool ret;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
*file = os_file_create_simple_no_error_handling(
|
|
|
|
0, path, OS_FILE_CREATE, OS_FILE_READ_WRITE, false, &ret);
|
2017-04-18 20:43:20 +02:00
|
|
|
if (!ret) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Can't create file %s", path);
|
2017-04-18 20:43:20 +02:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
ret = os_file_set_size(path, *file,
|
2018-04-27 13:26:43 +02:00
|
|
|
FIL_IBD_FILE_INITIAL_SIZE
|
|
|
|
<< srv_page_size_shift);
|
2017-04-18 20:43:20 +02:00
|
|
|
if (!ret) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: cannot set size for file %s", path);
|
2017-04-18 20:43:20 +02:00
|
|
|
os_file_close(*file);
|
|
|
|
os_file_delete(0, path);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
2018-03-27 10:34:43 +02:00
|
|
|
static fil_space_t* fil_space_get_by_name(const char* name)
|
|
|
|
{
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
mysql_mutex_assert_owner(&fil_system.mutex);
|
|
|
|
for (fil_space_t &space : fil_system.space_list)
|
|
|
|
if (space.chain.start)
|
|
|
|
if (const char *str= strstr(space.chain.start->name, name))
|
|
|
|
if (!strcmp(str + strlen(name), ".ibd") &&
|
|
|
|
(str == space.chain.start->name ||
|
|
|
|
IF_WIN(str[-1] == '\\' ||,) str[-1] == '/'))
|
|
|
|
return &space;
|
|
|
|
return nullptr;
|
2018-03-27 10:34:43 +02:00
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/***********************************************************************
|
|
|
|
Searches for matching tablespace file for given .delta file and space_id
|
|
|
|
in given directory. When matching tablespace found, renames it to match the
|
|
|
|
name of .delta file. If there was a tablespace with matching name and
|
|
|
|
mismatching ID, renames it to xtrabackup_tmp_#ID.ibd. If there was no
|
|
|
|
matching file, creates a new tablespace.
|
|
|
|
@return file handle of matched or created file */
|
|
|
|
static
|
2017-05-19 12:59:43 +02:00
|
|
|
pfs_os_file_t
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_delta_open_matching_space(
|
|
|
|
const char* dbname, /* in: path to destination database dir */
|
|
|
|
const char* name, /* in: name of delta file (without .delta) */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
const xb_delta_info_t& info,
|
2017-04-18 20:43:20 +02:00
|
|
|
char* real_name, /* out: full path of destination file */
|
|
|
|
size_t real_name_len, /* out: buffer size for real_name */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
bool* success) /* out: indicates error. true = success */
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
|
|
|
char dest_dir[FN_REFLEN];
|
|
|
|
char dest_space_name[FN_REFLEN];
|
|
|
|
fil_space_t* fil_space;
|
2017-05-19 12:59:43 +02:00
|
|
|
pfs_os_file_t file;
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_filter_entry_t* table;
|
|
|
|
|
|
|
|
ut_a(dbname != NULL ||
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
!fil_is_user_tablespace_id(info.space_id) ||
|
2021-07-22 10:22:47 +02:00
|
|
|
info.space_id == UINT32_MAX);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
*success = false;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (dbname) {
|
|
|
|
snprintf(dest_dir, FN_REFLEN, "%s/%s",
|
|
|
|
xtrabackup_target_dir, dbname);
|
|
|
|
snprintf(dest_space_name, FN_REFLEN, "%s/%s", dbname, name);
|
|
|
|
} else {
|
|
|
|
snprintf(dest_dir, FN_REFLEN, "%s", xtrabackup_target_dir);
|
|
|
|
snprintf(dest_space_name, FN_REFLEN, "%s", name);
|
|
|
|
}
|
|
|
|
|
|
|
|
snprintf(real_name, real_name_len,
|
|
|
|
"%s/%s",
|
|
|
|
xtrabackup_target_dir, dest_space_name);
|
|
|
|
/* Truncate ".ibd" */
|
|
|
|
dest_space_name[strlen(dest_space_name) - 4] = '\0';
|
|
|
|
|
|
|
|
/* Create the database directory if it doesn't exist yet */
|
|
|
|
if (!os_file_create_directory(dest_dir, FALSE)) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: error: cannot create dir %s", dest_dir);
|
2017-04-18 20:43:20 +02:00
|
|
|
return file;
|
|
|
|
}
|
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
if (!info.space_id && fil_system.sys_space) {
|
|
|
|
fil_node_t *node
|
|
|
|
= UT_LIST_GET_FIRST(fil_system.sys_space->chain);
|
|
|
|
for (; node; node = UT_LIST_GET_NEXT(chain, node)) {
|
|
|
|
if (!strcmp(node->name, real_name)) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (node && node->handle != OS_FILE_CLOSED) {
|
|
|
|
*success = true;
|
|
|
|
return node->handle;
|
|
|
|
}
|
|
|
|
msg("mariabackup: Cannot find file %s\n", real_name);
|
|
|
|
return OS_FILE_CLOSED;
|
|
|
|
}
|
|
|
|
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_lock(&recv_sys.mutex);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (!fil_is_user_tablespace_id(info.space_id)) {
|
|
|
|
found:
|
|
|
|
/* open the file and return its handle */
|
|
|
|
|
|
|
|
file = os_file_create_simple_no_error_handling(
|
|
|
|
0, real_name,
|
|
|
|
OS_FILE_OPEN, OS_FILE_READ_WRITE, false, success);
|
|
|
|
|
|
|
|
if (!*success) {
|
2017-10-25 16:17:21 +02:00
|
|
|
msg("mariabackup: Cannot open file %s\n", real_name);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
|
|
|
exit:
|
2022-02-10 15:37:12 +01:00
|
|
|
mysql_mutex_unlock(&recv_sys.mutex);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return file;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2021-06-14 08:15:20 +02:00
|
|
|
const size_t len = strlen(dest_space_name);
|
2017-04-18 20:43:20 +02:00
|
|
|
/* remember space name for further reference */
|
|
|
|
table = static_cast<xb_filter_entry_t *>
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
(malloc(sizeof(xb_filter_entry_t) +
|
2021-06-14 08:15:20 +02:00
|
|
|
len + 1));
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
table->name = ((char*)table) + sizeof(xb_filter_entry_t);
|
2021-06-14 08:15:20 +02:00
|
|
|
memcpy(table->name, dest_space_name, len + 1);
|
|
|
|
const ulint fold = my_crc32c(0, dest_space_name, len);
|
2020-06-18 11:26:28 +02:00
|
|
|
HASH_INSERT(xb_filter_entry_t, name_hash, &inc_dir_tables_hash,
|
2021-06-14 08:15:20 +02:00
|
|
|
fold, table);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_lock(&fil_system.mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
fil_space = fil_space_get_by_name(dest_space_name);
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_unlock(&fil_system.mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (fil_space != NULL) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (fil_space->id == info.space_id
|
2021-07-22 10:22:47 +02:00
|
|
|
|| info.space_id == UINT32_MAX) {
|
2017-04-18 20:43:20 +02:00
|
|
|
/* we found matching space */
|
|
|
|
goto found;
|
|
|
|
} else {
|
|
|
|
|
|
|
|
char tmpname[FN_REFLEN];
|
|
|
|
|
2021-07-22 10:22:47 +02:00
|
|
|
snprintf(tmpname, FN_REFLEN, "%s/xtrabackup_tmp_#%u",
|
2017-04-18 20:43:20 +02:00
|
|
|
dbname, fil_space->id);
|
|
|
|
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: Renaming %s to %s.ibd",
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
fil_space->chain.start->name, tmpname);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2021-05-14 07:24:43 +02:00
|
|
|
if (fil_space->rename(tmpname, false) != DB_SUCCESS) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: Cannot rename %s to %s",
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
fil_space->chain.start->name, tmpname);
|
2017-04-18 20:43:20 +02:00
|
|
|
goto exit;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2021-07-22 10:22:47 +02:00
|
|
|
if (info.space_id == UINT32_MAX)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2019-01-15 22:47:54 +01:00
|
|
|
die("Can't handle DDL operation on tablespace "
|
2017-04-18 20:43:20 +02:00
|
|
|
"%s\n", dest_space_name);
|
|
|
|
}
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_lock(&fil_system.mutex);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
fil_space = fil_space_get_by_id(info.space_id);
|
2020-12-04 18:02:58 +01:00
|
|
|
mysql_mutex_unlock(&fil_system.mutex);
|
2017-04-18 20:43:20 +02:00
|
|
|
if (fil_space != NULL) {
|
|
|
|
char tmpname[FN_REFLEN];
|
|
|
|
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
snprintf(tmpname, sizeof tmpname, "%s.ibd", dest_space_name);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: Renaming %s to %s",
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
fil_space->chain.start->name, tmpname);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
if (fil_space->rename(tmpname, false) != DB_SUCCESS) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: Cannot rename %s to %s",
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
fil_space->chain.start->name, tmpname);
|
2017-04-18 20:43:20 +02:00
|
|
|
goto exit;
|
|
|
|
}
|
|
|
|
|
|
|
|
goto found;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* No matching space found. create the new one. */
|
2021-07-22 10:22:47 +02:00
|
|
|
const uint32_t flags = info.zip_size
|
2019-02-06 18:50:11 +01:00
|
|
|
? get_bit_shift(info.page_size
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
>> (UNIV_ZIP_SIZE_SHIFT_MIN - 1))
|
|
|
|
<< FSP_FLAGS_POS_ZIP_SSIZE
|
|
|
|
| FSP_FLAGS_MASK_POST_ANTELOPE
|
|
|
|
| FSP_FLAGS_MASK_ATOMIC_BLOBS
|
2019-02-06 18:50:11 +01:00
|
|
|
| (srv_page_size == UNIV_PAGE_SIZE_ORIG
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
? 0
|
2019-02-06 18:50:11 +01:00
|
|
|
: get_bit_shift(srv_page_size
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
>> (UNIV_ZIP_SIZE_SHIFT_MIN - 1))
|
|
|
|
<< FSP_FLAGS_POS_PAGE_SSIZE)
|
|
|
|
: FSP_FLAGS_PAGE_SSIZE();
|
2019-02-06 18:50:11 +01:00
|
|
|
ut_ad(fil_space_t::zip_size(flags) == info.zip_size);
|
|
|
|
ut_ad(fil_space_t::physical_size(flags) == info.page_size);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
if (fil_space_t::create(info.space_id, flags,
|
2020-10-26 14:59:30 +01:00
|
|
|
FIL_TYPE_TABLESPACE, 0)) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
*success = xb_space_create_file(real_name, info.space_id,
|
|
|
|
flags, &file);
|
|
|
|
} else {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Can't create tablespace %s\n", dest_space_name);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
goto exit;
|
|
|
|
}
|
|
|
|
|
|
|
|
/************************************************************************
|
|
|
|
Applies a given .delta file to the corresponding data file.
|
|
|
|
@return TRUE on success */
|
|
|
|
static
|
|
|
|
ibool
|
|
|
|
xtrabackup_apply_delta(
|
|
|
|
const char* dirname, /* in: dir name of incremental */
|
|
|
|
const char* dbname, /* in: database name (ibdata: NULL) */
|
|
|
|
const char* filename, /* in: file name (not a path),
|
|
|
|
including the .delta extension */
|
|
|
|
void* /*data*/)
|
|
|
|
{
|
2017-05-19 12:59:43 +02:00
|
|
|
pfs_os_file_t src_file;
|
|
|
|
pfs_os_file_t dst_file;
|
2017-04-18 20:43:20 +02:00
|
|
|
char src_path[FN_REFLEN];
|
|
|
|
char dst_path[FN_REFLEN];
|
|
|
|
char meta_path[FN_REFLEN];
|
|
|
|
char space_name[FN_REFLEN];
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
bool success;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
ibool last_buffer = FALSE;
|
|
|
|
ulint page_in_buffer;
|
|
|
|
ulint incremental_buffers = 0;
|
|
|
|
|
2019-02-06 18:50:11 +01:00
|
|
|
xb_delta_info_t info(srv_page_size, 0, SRV_TMP_SPACE_ID);
|
2017-04-18 20:43:20 +02:00
|
|
|
ulint page_size;
|
|
|
|
ulint page_size_shift;
|
2019-12-05 05:42:31 +01:00
|
|
|
byte* incremental_buffer = NULL;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
size_t offset;
|
|
|
|
|
|
|
|
ut_a(xtrabackup_incremental);
|
|
|
|
|
|
|
|
if (dbname) {
|
|
|
|
snprintf(src_path, sizeof(src_path), "%s/%s/%s",
|
|
|
|
dirname, dbname, filename);
|
|
|
|
snprintf(dst_path, sizeof(dst_path), "%s/%s/%s",
|
|
|
|
xtrabackup_real_target_dir, dbname, filename);
|
|
|
|
} else {
|
|
|
|
snprintf(src_path, sizeof(src_path), "%s/%s",
|
|
|
|
dirname, filename);
|
|
|
|
snprintf(dst_path, sizeof(dst_path), "%s/%s",
|
|
|
|
xtrabackup_real_target_dir, filename);
|
|
|
|
}
|
|
|
|
dst_path[strlen(dst_path) - 6] = '\0';
|
|
|
|
|
2019-03-06 10:22:27 +01:00
|
|
|
strncpy(space_name, filename, FN_REFLEN - 1);
|
|
|
|
space_name[FN_REFLEN - 1] = '\0';
|
2017-04-18 20:43:20 +02:00
|
|
|
space_name[strlen(space_name) - 6] = 0;
|
|
|
|
|
|
|
|
if (!get_meta_path(src_path, meta_path)) {
|
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!xb_read_delta_metadata(meta_path, &info)) {
|
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
2019-02-06 18:50:11 +01:00
|
|
|
page_size = info.page_size;
|
2017-04-18 20:43:20 +02:00
|
|
|
page_size_shift = get_bit_shift(page_size);
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("page size for %s is %zu bytes",
|
2017-04-18 20:43:20 +02:00
|
|
|
src_path, page_size);
|
|
|
|
if (page_size_shift < 10 ||
|
|
|
|
page_size_shift > UNIV_PAGE_SIZE_SHIFT_MAX) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("error: invalid value of page_size "
|
|
|
|
"(%zu bytes) read from %s", page_size, meta_path);
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
src_file = os_file_create_simple_no_error_handling(
|
|
|
|
0, src_path,
|
|
|
|
OS_FILE_OPEN, OS_FILE_READ_WRITE, false, &success);
|
2017-04-18 20:43:20 +02:00
|
|
|
if (!success) {
|
|
|
|
os_file_get_last_error(TRUE);
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("error: can't open %s", src_path);
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
|
|
|
posix_fadvise(src_file, 0, 0, POSIX_FADV_SEQUENTIAL);
|
|
|
|
|
|
|
|
dst_file = xb_delta_open_matching_space(
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
dbname, space_name, info,
|
2017-04-18 20:43:20 +02:00
|
|
|
dst_path, sizeof(dst_path), &success);
|
|
|
|
if (!success) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("error: can't open %s", dst_path);
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
|
|
|
posix_fadvise(dst_file, 0, 0, POSIX_FADV_DONTNEED);
|
|
|
|
|
|
|
|
/* allocate buffer for incremental backup (4096 pages) */
|
|
|
|
incremental_buffer = static_cast<byte *>
|
2019-12-05 05:42:31 +01:00
|
|
|
(aligned_malloc(page_size / 4 * page_size, page_size));
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Applying %s to %s...", src_path, dst_path);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
while (!last_buffer) {
|
|
|
|
ulint cluster_header;
|
|
|
|
|
|
|
|
/* read to buffer */
|
|
|
|
/* first block of block cluster */
|
|
|
|
offset = ((incremental_buffers * (page_size / 4))
|
|
|
|
<< page_size_shift);
|
2019-05-05 14:03:48 +02:00
|
|
|
if (os_file_read(IORequestRead, src_file,
|
2022-11-30 09:54:03 +01:00
|
|
|
incremental_buffer, offset, page_size,
|
|
|
|
nullptr)
|
2019-05-05 14:03:48 +02:00
|
|
|
!= DB_SUCCESS) {
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
|
|
|
cluster_header = mach_read_from_4(incremental_buffer);
|
|
|
|
switch(cluster_header) {
|
|
|
|
case 0x78747261UL: /*"xtra"*/
|
|
|
|
break;
|
|
|
|
case 0x58545241UL: /*"XTRA"*/
|
|
|
|
last_buffer = TRUE;
|
|
|
|
break;
|
|
|
|
default:
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("error: %s seems not "
|
|
|
|
".delta file.", src_path);
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/* FIXME: If the .delta modifies FSP_SIZE on page 0,
|
|
|
|
extend the file to that size. */
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
for (page_in_buffer = 1; page_in_buffer < page_size / 4;
|
|
|
|
page_in_buffer++) {
|
|
|
|
if (mach_read_from_4(incremental_buffer + page_in_buffer * 4)
|
|
|
|
== 0xFFFFFFFFUL)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
ut_a(last_buffer || page_in_buffer == page_size / 4);
|
|
|
|
|
|
|
|
/* read whole of the cluster */
|
2019-05-05 14:03:48 +02:00
|
|
|
if (os_file_read(IORequestRead, src_file,
|
|
|
|
incremental_buffer,
|
2022-11-30 09:54:03 +01:00
|
|
|
offset, page_in_buffer * page_size, nullptr)
|
2019-05-05 14:03:48 +02:00
|
|
|
!= DB_SUCCESS) {
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
|
|
|
posix_fadvise(src_file, offset, page_in_buffer * page_size,
|
|
|
|
POSIX_FADV_DONTNEED);
|
|
|
|
|
|
|
|
for (page_in_buffer = 1; page_in_buffer < page_size / 4;
|
|
|
|
page_in_buffer++) {
|
|
|
|
ulint offset_on_page;
|
|
|
|
|
|
|
|
offset_on_page = mach_read_from_4(incremental_buffer + page_in_buffer * 4);
|
|
|
|
|
|
|
|
if (offset_on_page == 0xFFFFFFFFUL)
|
|
|
|
break;
|
|
|
|
|
2017-10-09 00:05:54 +02:00
|
|
|
uchar *buf = incremental_buffer + page_in_buffer * page_size;
|
|
|
|
const os_offset_t off = os_offset_t(offset_on_page)*page_size;
|
|
|
|
|
|
|
|
if (off == 0) {
|
|
|
|
/* Read tablespace size from page 0,
|
|
|
|
and extend the file to specified size.*/
|
2017-11-24 16:17:16 +01:00
|
|
|
os_offset_t n_pages = mach_read_from_4(
|
|
|
|
buf + FSP_HEADER_OFFSET + FSP_SIZE);
|
|
|
|
if (mach_read_from_4(buf
|
|
|
|
+ FIL_PAGE_SPACE_ID)) {
|
|
|
|
if (!os_file_set_size(
|
|
|
|
dst_path, dst_file,
|
|
|
|
n_pages * page_size))
|
2017-11-23 22:01:00 +01:00
|
|
|
goto error;
|
2017-11-24 16:17:16 +01:00
|
|
|
} else if (fil_space_t* space
|
2018-04-23 12:14:28 +02:00
|
|
|
= fil_system.sys_space) {
|
2017-11-24 16:17:16 +01:00
|
|
|
/* The system tablespace can
|
|
|
|
consist of multiple files. The
|
|
|
|
first one has full tablespace
|
|
|
|
size in page 0, but only the last
|
|
|
|
file should be extended. */
|
|
|
|
fil_node_t* n = UT_LIST_GET_FIRST(
|
|
|
|
space->chain);
|
|
|
|
bool fail = !strcmp(n->name, dst_path)
|
|
|
|
&& !fil_space_extend(
|
2020-10-15 15:28:19 +02:00
|
|
|
space, uint32_t(n_pages));
|
2017-11-24 16:17:16 +01:00
|
|
|
if (fail) goto error;
|
2017-11-23 22:01:00 +01:00
|
|
|
}
|
2017-10-09 00:05:54 +02:00
|
|
|
}
|
|
|
|
|
2019-05-05 14:03:48 +02:00
|
|
|
if (os_file_write(IORequestWrite,
|
|
|
|
dst_path, dst_file, buf, off,
|
|
|
|
page_size) != DB_SUCCESS) {
|
2017-04-18 20:43:20 +02:00
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-04-12 13:09:32 +02:00
|
|
|
/* Free file system buffer cache after the batch was written. */
|
|
|
|
#ifdef __linux__
|
|
|
|
os_file_flush_func(dst_file);
|
|
|
|
#endif
|
|
|
|
posix_fadvise(dst_file, 0, 0, POSIX_FADV_DONTNEED);
|
|
|
|
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
incremental_buffers++;
|
|
|
|
}
|
|
|
|
|
2019-12-05 05:42:31 +01:00
|
|
|
aligned_free(incremental_buffer);
|
2017-10-09 00:05:54 +02:00
|
|
|
if (src_file != OS_FILE_CLOSED) {
|
2017-04-18 20:43:20 +02:00
|
|
|
os_file_close(src_file);
|
2017-10-09 00:05:54 +02:00
|
|
|
os_file_delete(0,src_path);
|
|
|
|
}
|
2020-10-26 14:59:30 +01:00
|
|
|
if (dst_file != OS_FILE_CLOSED && info.space_id)
|
2017-04-18 20:43:20 +02:00
|
|
|
os_file_close(dst_file);
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
error:
|
2019-12-05 05:42:31 +01:00
|
|
|
aligned_free(incremental_buffer);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (src_file != OS_FILE_CLOSED)
|
2017-04-18 20:43:20 +02:00
|
|
|
os_file_close(src_file);
|
2020-10-26 14:59:30 +01:00
|
|
|
if (dst_file != OS_FILE_CLOSED && info.space_id)
|
2017-04-18 20:43:20 +02:00
|
|
|
os_file_close(dst_file);
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: xtrabackup_apply_delta(): "
|
2017-04-18 20:43:20 +02:00
|
|
|
"failed to apply %s to %s.\n", src_path, dst_path);
|
|
|
|
return FALSE;
|
|
|
|
}
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
|
|
|
|
std::string change_extension(std::string filename, std::string new_ext) {
|
|
|
|
DBUG_ASSERT(new_ext.size() == 3);
|
|
|
|
std::string new_name(filename);
|
|
|
|
new_name.resize(new_name.size() - new_ext.size());
|
|
|
|
new_name.append(new_ext);
|
|
|
|
return new_name;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static void rename_file(const char *from,const char *to) {
|
|
|
|
msg("Renaming %s to %s\n", from, to);
|
|
|
|
if (my_rename(from, to, MY_WME)) {
|
2019-01-15 22:47:54 +01:00
|
|
|
die("Can't rename %s to %s errno %d", from, to, errno);
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void rename_file(const std::string& from, const std::string &to) {
|
|
|
|
rename_file(from.c_str(), to.c_str());
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
/************************************************************************
|
|
|
|
Callback to handle datadir entry. Function of this type will be called
|
|
|
|
for each entry which matches the mask by xb_process_datadir.
|
|
|
|
@return should return TRUE on success */
|
|
|
|
typedef ibool (*handle_datadir_entry_func_t)(
|
|
|
|
/*=========================================*/
|
|
|
|
const char* data_home_dir, /*!<in: path to datadir */
|
|
|
|
const char* db_name, /*!<in: database name */
|
|
|
|
const char* file_name, /*!<in: file name with suffix */
|
|
|
|
void* arg); /*!<in: caller-provided data */
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
/** Rename, and replace destination file, if exists */
|
|
|
|
static void rename_force(const char *from, const char *to) {
|
|
|
|
if (access(to, R_OK) == 0) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Removing %s", to);
|
2018-08-09 16:06:52 +02:00
|
|
|
if (my_delete(to, MYF(MY_WME))) {
|
|
|
|
msg("Can't remove %s, errno %d", to, errno);
|
|
|
|
exit(EXIT_FAILURE);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
rename_file(from,to);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
/** During prepare phase, rename ".new" files, that were created in
|
|
|
|
backup_fix_ddl() and backup_optimized_ddl_op(), to ".ibd". In the case of
|
|
|
|
incremental backup, i.e. of arg argument is set, move ".new" files to
|
|
|
|
destination directory and rename them to ".ibd", remove existing ".ibd.delta"
|
|
|
|
and ".idb.meta" files in incremental directory to avoid applying delta to
|
|
|
|
".ibd" file.
|
|
|
|
|
|
|
|
@param[in] data_home_dir path to datadir
|
|
|
|
@param[in] db_name database name
|
|
|
|
@param[in] file_name file name with suffix
|
|
|
|
@param[in] arg destination path, used in incremental backup to notify, that
|
|
|
|
*.new file must be moved to destibation directory
|
|
|
|
|
|
|
|
@return true */
|
|
|
|
static ibool prepare_handle_new_files(const char *data_home_dir,
|
|
|
|
const char *db_name,
|
|
|
|
const char *file_name, void *arg)
|
|
|
|
{
|
|
|
|
const char *dest_dir = static_cast<const char *>(arg);
|
2022-04-30 06:49:13 +02:00
|
|
|
std::string src_path = std::string(data_home_dir) + '/' + std::string(db_name) + '/';
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
/* Copy "*.new" files from incremental to base dir for incremental backup */
|
|
|
|
std::string dest_path=
|
|
|
|
dest_dir ? std::string(dest_dir) + '/' + std::string(db_name) +
|
2022-04-30 06:49:13 +02:00
|
|
|
'/' : src_path;
|
|
|
|
|
|
|
|
/*
|
|
|
|
A CREATE DATABASE could have happened during the base mariabackup run.
|
|
|
|
In case if the current table file (e.g. `t1.new`) is from such
|
|
|
|
a new database, the database directory may not exist yet in
|
|
|
|
the base backup directory. Let's make sure to check if the directory
|
|
|
|
exists (and create if needed).
|
|
|
|
*/
|
|
|
|
if (!directory_exists(dest_path.c_str(), true/*create if not exists*/))
|
|
|
|
return FALSE;
|
|
|
|
src_path+= file_name;
|
|
|
|
dest_path+= file_name;
|
2018-08-09 16:06:52 +02:00
|
|
|
|
|
|
|
size_t index = dest_path.find(".new");
|
|
|
|
DBUG_ASSERT(index != std::string::npos);
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
dest_path.replace(index, strlen(".ibd"), ".ibd");
|
2018-08-09 16:06:52 +02:00
|
|
|
rename_force(src_path.c_str(),dest_path.c_str());
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
|
|
|
|
if (dest_dir) {
|
|
|
|
/* remove delta and meta files to avoid delta applying for new file */
|
|
|
|
index = src_path.find(".new");
|
|
|
|
DBUG_ASSERT(index != std::string::npos);
|
|
|
|
src_path.replace(index, std::string::npos, ".ibd.delta");
|
|
|
|
if (access(src_path.c_str(), R_OK) == 0) {
|
|
|
|
msg("Removing %s", src_path.c_str());
|
|
|
|
if (my_delete(src_path.c_str(), MYF(MY_WME)))
|
|
|
|
die("Can't remove %s, errno %d", src_path.c_str(), errno);
|
|
|
|
}
|
|
|
|
src_path.replace(index, std::string::npos, ".ibd.meta");
|
|
|
|
if (access(src_path.c_str(), R_OK) == 0) {
|
|
|
|
msg("Removing %s", src_path.c_str());
|
|
|
|
if (my_delete(src_path.c_str(), MYF(MY_WME)))
|
|
|
|
die("Can't remove %s, errno %d", src_path.c_str(), errno);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* add table name to the container to avoid it's deletion at the end of
|
|
|
|
prepare */
|
|
|
|
std::string table_name = std::string(db_name) + '/'
|
|
|
|
+ std::string(file_name, file_name + strlen(file_name) - strlen(".new"));
|
|
|
|
xb_filter_entry_t *table = static_cast<xb_filter_entry_t *>
|
|
|
|
(malloc(sizeof(xb_filter_entry_t) + table_name.size() + 1));
|
|
|
|
table->name = ((char*)table) + sizeof(xb_filter_entry_t);
|
|
|
|
strcpy(table->name, table_name.c_str());
|
2021-06-14 08:15:20 +02:00
|
|
|
const ulint fold = my_crc32c(0, table->name,
|
|
|
|
table_name.size());
|
2020-10-30 10:15:30 +01:00
|
|
|
HASH_INSERT(xb_filter_entry_t, name_hash, &inc_dir_tables_hash,
|
2021-06-14 08:15:20 +02:00
|
|
|
fold, table);
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
}
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/************************************************************************
|
|
|
|
Callback to handle datadir entry. Deletes entry if it has no matching
|
|
|
|
fil_space in fil_system directory.
|
|
|
|
@return FALSE if delete attempt was unsuccessful */
|
|
|
|
static
|
|
|
|
ibool
|
|
|
|
rm_if_not_found(
|
|
|
|
const char* data_home_dir, /*!<in: path to datadir */
|
|
|
|
const char* db_name, /*!<in: database name */
|
|
|
|
const char* file_name, /*!<in: file name with suffix */
|
|
|
|
void* arg __attribute__((unused)))
|
|
|
|
{
|
|
|
|
char name[FN_REFLEN];
|
|
|
|
xb_filter_entry_t* table;
|
|
|
|
|
|
|
|
snprintf(name, FN_REFLEN, "%s/%s", db_name, file_name);
|
|
|
|
/* Truncate ".ibd" */
|
2021-06-14 08:15:20 +02:00
|
|
|
const size_t len = strlen(name) - 4;
|
|
|
|
name[len] = '\0';
|
|
|
|
const ulint fold = my_crc32c(0, name, len);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2021-06-14 08:15:20 +02:00
|
|
|
HASH_SEARCH(name_hash, &inc_dir_tables_hash, fold,
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_filter_entry_t*,
|
|
|
|
table, (void) 0,
|
|
|
|
!strcmp(table->name, name));
|
|
|
|
|
|
|
|
if (!table) {
|
|
|
|
snprintf(name, FN_REFLEN, "%s/%s/%s", data_home_dir,
|
|
|
|
db_name, file_name);
|
|
|
|
return os_file_delete(0, name);
|
|
|
|
}
|
|
|
|
|
|
|
|
return(TRUE);
|
|
|
|
}
|
|
|
|
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
/** Function enumerates files in datadir (provided by path) which are matched
|
2017-04-18 20:43:20 +02:00
|
|
|
by provided suffix. For each entry callback is called.
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
|
|
|
|
@param[in] path datadir path
|
|
|
|
@param[in] suffix suffix to match against
|
|
|
|
@param[in] func callback
|
|
|
|
@param[in] func_arg arguments for the above callback
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
@return FALSE if callback for some entry returned FALSE */
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
static ibool xb_process_datadir(const char *path, const char *suffix,
|
|
|
|
handle_datadir_entry_func_t func,
|
|
|
|
void *func_arg = NULL)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
|
|
|
ulint ret;
|
2019-04-19 12:15:46 +02:00
|
|
|
char dbpath[OS_FILE_MAX_PATH+2];
|
2017-04-18 20:43:20 +02:00
|
|
|
os_file_dir_t dir;
|
|
|
|
os_file_dir_t dbdir;
|
|
|
|
os_file_stat_t dbinfo;
|
|
|
|
os_file_stat_t fileinfo;
|
|
|
|
ulint suffix_len;
|
|
|
|
dberr_t err = DB_SUCCESS;
|
|
|
|
static char current_dir[2];
|
|
|
|
|
|
|
|
current_dir[0] = FN_CURLIB;
|
|
|
|
current_dir[1] = 0;
|
|
|
|
srv_data_home = current_dir;
|
|
|
|
|
|
|
|
suffix_len = strlen(suffix);
|
|
|
|
|
|
|
|
/* datafile */
|
2021-05-18 11:13:18 +02:00
|
|
|
dbdir = os_file_opendir(path);
|
|
|
|
if (UNIV_UNLIKELY(dbdir != IF_WIN(INVALID_HANDLE_VALUE, nullptr))) {
|
|
|
|
ret = fil_file_readdir_next_file(&err, path, dbdir, &fileinfo);
|
2017-04-18 20:43:20 +02:00
|
|
|
while (ret == 0) {
|
|
|
|
if (fileinfo.type == OS_FILE_TYPE_DIR) {
|
|
|
|
goto next_file_item_1;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (strlen(fileinfo.name) > suffix_len
|
|
|
|
&& 0 == strcmp(fileinfo.name +
|
|
|
|
strlen(fileinfo.name) - suffix_len,
|
|
|
|
suffix)) {
|
|
|
|
if (!func(
|
|
|
|
path, NULL,
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
fileinfo.name, func_arg))
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2018-12-03 12:59:45 +01:00
|
|
|
os_file_closedir(dbdir);
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
next_file_item_1:
|
|
|
|
ret = fil_file_readdir_next_file(&err,
|
|
|
|
path, dbdir,
|
|
|
|
&fileinfo);
|
|
|
|
}
|
|
|
|
|
|
|
|
os_file_closedir(dbdir);
|
|
|
|
} else {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Can't open dir %s", path);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* single table tablespaces */
|
2021-05-18 11:13:18 +02:00
|
|
|
dir = os_file_opendir(path);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
if (UNIV_UNLIKELY(dbdir == IF_WIN(INVALID_HANDLE_VALUE, nullptr))) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Can't open dir %s", path);
|
2021-05-18 11:13:18 +02:00
|
|
|
return TRUE;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
ret = fil_file_readdir_next_file(&err, path, dir, &dbinfo);
|
2017-04-18 20:43:20 +02:00
|
|
|
while (ret == 0) {
|
|
|
|
if (dbinfo.type == OS_FILE_TYPE_FILE
|
|
|
|
|| dbinfo.type == OS_FILE_TYPE_UNKNOWN) {
|
|
|
|
|
|
|
|
goto next_datadir_item;
|
|
|
|
}
|
|
|
|
|
2019-06-02 12:53:16 +02:00
|
|
|
snprintf(dbpath, sizeof(dbpath), "%.*s/%.*s",
|
|
|
|
OS_FILE_MAX_PATH/2-1,
|
|
|
|
path,
|
|
|
|
OS_FILE_MAX_PATH/2-1,
|
|
|
|
dbinfo.name);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
dbdir = os_file_opendir(dbpath);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2021-05-18 11:13:18 +02:00
|
|
|
if (dbdir != IF_WIN(INVALID_HANDLE_VALUE, nullptr)) {
|
2017-04-18 20:43:20 +02:00
|
|
|
ret = fil_file_readdir_next_file(&err, dbpath, dbdir,
|
|
|
|
&fileinfo);
|
|
|
|
while (ret == 0) {
|
|
|
|
|
|
|
|
if (fileinfo.type == OS_FILE_TYPE_DIR) {
|
|
|
|
|
|
|
|
goto next_file_item_2;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (strlen(fileinfo.name) > suffix_len
|
|
|
|
&& 0 == strcmp(fileinfo.name +
|
|
|
|
strlen(fileinfo.name) -
|
|
|
|
suffix_len,
|
|
|
|
suffix)) {
|
|
|
|
/* The name ends in suffix; process
|
|
|
|
the file */
|
|
|
|
if (!func(
|
|
|
|
path,
|
|
|
|
dbinfo.name,
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
fileinfo.name, func_arg))
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2018-12-03 12:59:45 +01:00
|
|
|
os_file_closedir(dbdir);
|
2018-12-04 14:45:27 +01:00
|
|
|
os_file_closedir(dir);
|
2017-04-18 20:43:20 +02:00
|
|
|
return(FALSE);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
next_file_item_2:
|
|
|
|
ret = fil_file_readdir_next_file(&err,
|
|
|
|
dbpath, dbdir,
|
|
|
|
&fileinfo);
|
|
|
|
}
|
|
|
|
|
|
|
|
os_file_closedir(dbdir);
|
|
|
|
}
|
|
|
|
next_datadir_item:
|
|
|
|
ret = fil_file_readdir_next_file(&err,
|
|
|
|
path,
|
|
|
|
dir, &dbinfo);
|
|
|
|
}
|
|
|
|
|
|
|
|
os_file_closedir(dir);
|
|
|
|
|
|
|
|
return(TRUE);
|
|
|
|
}
|
|
|
|
|
|
|
|
/************************************************************************
|
|
|
|
Applies all .delta files from incremental_dir to the full backup.
|
|
|
|
@return TRUE on success. */
|
|
|
|
static
|
|
|
|
ibool
|
|
|
|
xtrabackup_apply_deltas()
|
|
|
|
{
|
|
|
|
return xb_process_datadir(xtrabackup_incremental_dir, ".delta",
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
xtrabackup_apply_delta);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static
|
|
|
|
void
|
|
|
|
innodb_free_param()
|
|
|
|
{
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
srv_sys_space.shutdown();
|
2017-04-18 20:43:20 +02:00
|
|
|
free_tmpdir(&mysql_tmpdir_list);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
/** Check if file exists*/
|
|
|
|
static bool file_exists(std::string name)
|
|
|
|
{
|
|
|
|
return access(name.c_str(), R_OK) == 0 ;
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Read file content into STL string */
|
|
|
|
static std::string read_file_as_string(const std::string file) {
|
|
|
|
char content[FN_REFLEN];
|
|
|
|
FILE *f = fopen(file.c_str(), "r");
|
|
|
|
if (!f) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Can not open %s", file.c_str());
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
size_t len = fread(content, 1, FN_REFLEN, f);
|
|
|
|
fclose(f);
|
|
|
|
return std::string(content, len);
|
|
|
|
}
|
|
|
|
|
|
|
|
/** Delete file- Provide verbose diagnostics and exit, if operation fails. */
|
|
|
|
static void delete_file(const std::string& file, bool if_exists = false) {
|
|
|
|
if (if_exists && !file_exists(file))
|
|
|
|
return;
|
|
|
|
if (my_delete(file.c_str(), MYF(MY_WME))) {
|
2019-01-15 22:47:54 +01:00
|
|
|
die("Can't remove %s, errno %d", file.c_str(), errno);
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
Rename tablespace during prepare.
|
|
|
|
Backup in its end phase may generate some .ren files, recording
|
|
|
|
tablespaces that should be renamed in --prepare.
|
|
|
|
*/
|
|
|
|
static void rename_table_in_prepare(const std::string &datadir, const std::string& from , const std::string& to,
|
|
|
|
const char *extension=0) {
|
|
|
|
if (!extension) {
|
|
|
|
static const char *extensions_nonincremental[] = { ".ibd", 0 };
|
|
|
|
static const char *extensions_incremental[] = { ".ibd.delta", ".ibd.meta", 0 };
|
|
|
|
const char **extensions = xtrabackup_incremental_dir ?
|
|
|
|
extensions_incremental : extensions_nonincremental;
|
|
|
|
for (size_t i = 0; extensions[i]; i++) {
|
|
|
|
rename_table_in_prepare(datadir, from, to, extensions[i]);
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
std::string src = std::string(datadir) + "/" + from + extension;
|
|
|
|
std::string dest = std::string(datadir) + "/" + to + extension;
|
|
|
|
std::string ren2, tmp;
|
|
|
|
if (file_exists(dest)) {
|
|
|
|
ren2= std::string(datadir) + "/" + to + ".ren";
|
|
|
|
if (!file_exists(ren2)) {
|
MDEV-25312 Replace fil_space_t::name with fil_space_t::name()
A consistency check for fil_space_t::name is causing recovery failures
in MDEV-25180 (Atomic ALTER TABLE). So, we'd better remove that field
altogether.
fil_space_t::name was more or less a copy of dict_table_t::name
(except for some special cases), and it was not being used for
anything useful.
There used to be a name_hash, but it had been removed already in
commit a75dbfd7183cc96680f3e3e684fd36500dac8158 (MDEV-12266).
We will also remove os_normalize_path(), OS_PATH_SEPARATOR,
OS_PATH_SEPATOR_ALT. On Microsoft Windows, we will treat \ and /
roughly in the same way. The intention is that for per-table
tablespaces, the filenames will always follow the pattern
prefix/databasename/tablename.ibd. (Any \ in the prefix must not
be converted.)
ut_basename_noext(): Remove (unused function).
read_link_file(): Replaces RemoteDatafile::read_link_file().
We will ensure that the last two path component separators are
forward slashes (converting up to 2 trailing backslashes on
Microsoft Windows), so that everywhere else we can
assume that data file names end in "/databasename/tablename.ibd".
Note: On Microsoft Windows, path names that start with \\?\ must
not contain / as path component separators. Previously, such paths
did work in the DATA DIRECTORY argument of InnoDB tables.
Reviewed by: Vladislav Vaintroub
2021-04-07 17:01:13 +02:00
|
|
|
die("ERROR : File %s was not found, but expected during rename processing\n", ren2.c_str());
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
tmp = to + "#";
|
|
|
|
rename_table_in_prepare(datadir, to, tmp);
|
|
|
|
}
|
|
|
|
rename_file(src, dest);
|
|
|
|
if (ren2.size()) {
|
|
|
|
// Make sure the temp. renamed file is processed.
|
|
|
|
std::string to2 = read_file_as_string(ren2);
|
|
|
|
rename_table_in_prepare(datadir, tmp, to2);
|
|
|
|
delete_file(ren2);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static ibool prepare_handle_ren_files(const char *datadir, const char *db, const char *filename, void *) {
|
|
|
|
|
|
|
|
std::string ren_file = std::string(datadir) + "/" + db + "/" + filename;
|
|
|
|
if (!file_exists(ren_file))
|
|
|
|
return TRUE;
|
|
|
|
|
|
|
|
std::string to = read_file_as_string(ren_file);
|
|
|
|
std::string source_space_name = std::string(db) + "/" + filename;
|
|
|
|
source_space_name.resize(source_space_name.size() - 4); // remove extension
|
|
|
|
|
|
|
|
rename_table_in_prepare(datadir, source_space_name.c_str(), to.c_str());
|
|
|
|
delete_file(ren_file);
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Remove tablespaces during backup, based on */
|
|
|
|
static ibool prepare_handle_del_files(const char *datadir, const char *db, const char *filename, void *) {
|
|
|
|
std::string del_file = std::string(datadir) + "/" + db + "/" + filename;
|
|
|
|
std::string path(del_file);
|
|
|
|
path.resize(path.size() - 4); // remove extension;
|
|
|
|
if (xtrabackup_incremental) {
|
|
|
|
delete_file(path + ".ibd.delta", true);
|
|
|
|
delete_file(path + ".ibd.meta", true);
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
delete_file(path + ".ibd", true);
|
|
|
|
}
|
|
|
|
delete_file(del_file);
|
|
|
|
return TRUE;
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/** Implement --prepare
|
|
|
|
@return whether the operation succeeded */
|
2019-03-22 10:15:21 +01:00
|
|
|
static bool xtrabackup_prepare_func(char** argv)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
CorruptedPages corrupted_pages;
|
2017-04-18 20:43:20 +02:00
|
|
|
char metadata_path[FN_REFLEN];
|
|
|
|
|
|
|
|
/* cd to target-dir */
|
|
|
|
|
|
|
|
if (my_setwd(xtrabackup_real_target_dir,MYF(MY_WME)))
|
|
|
|
{
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("can't my_setwd %s", xtrabackup_real_target_dir);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(false);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("cd to %s", xtrabackup_real_target_dir);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2018-08-09 16:06:52 +02:00
|
|
|
fil_path_to_mysql_datadir = ".";
|
|
|
|
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
ut_ad(xtrabackup_incremental == xtrabackup_incremental_dir);
|
2020-10-30 10:15:30 +01:00
|
|
|
if (xtrabackup_incremental)
|
|
|
|
inc_dir_tables_hash.create(1000);
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
|
2021-01-06 10:53:00 +01:00
|
|
|
msg("open files limit requested %u, set to %lu",
|
2020-12-15 14:46:58 +01:00
|
|
|
(uint) xb_open_files_limit,
|
|
|
|
xb_set_max_open_files(xb_open_files_limit));
|
|
|
|
|
2019-01-10 19:35:45 +01:00
|
|
|
/* Fix DDL for prepare. Process .del,.ren, and .new files.
|
|
|
|
The order in which files are processed, is important
|
|
|
|
(see MDEV-18185, MDEV-18201)
|
|
|
|
*/
|
|
|
|
xb_process_datadir(xtrabackup_incremental_dir ? xtrabackup_incremental_dir : ".",
|
|
|
|
".del", prepare_handle_del_files);
|
|
|
|
xb_process_datadir(xtrabackup_incremental_dir? xtrabackup_incremental_dir:".",
|
|
|
|
".ren", prepare_handle_ren_files);
|
2018-08-09 16:06:52 +02:00
|
|
|
if (xtrabackup_incremental_dir) {
|
|
|
|
xb_process_datadir(xtrabackup_incremental_dir, ".new.meta", prepare_handle_new_files);
|
|
|
|
xb_process_datadir(xtrabackup_incremental_dir, ".new.delta", prepare_handle_new_files);
|
MDEV-20755 InnoDB: Database page corruption on disk or a failed file read of tablespace upon prepare of mariabackup incremental backup
The problem:
When incremental backup is taken, delta files are created for innodb tables
which are marked as new tables during innodb ddl tracking. When such
tablespace is tried to be opened during prepare in
xb_delta_open_matching_space(), it is "created", i.e.
xb_space_create_file() is invoked, instead of opening, even if
a tablespace with the same name exists in the base backup directory.
xb_space_create_file() writes page 0 header the tablespace.
This header does not contain crypt data, as mariabackup does not have
any information about crypt data in delta file metadata for
tablespaces.
After delta file is applied, recovery process is started. As the
sequence of recovery for different pages is not defined, there can be
the situation when crypt data redo log event is executed after some
other page is read for recovery. When some page is read for recovery, it's
decrypted using crypt data stored in tablespace header in page 0, if
there is no crypt data, the page is not decryped and does not pass corruption
test.
This causes error for incremental backup --prepare for encrypted
tablespaces.
The error is not stable because crypt data redo log event updates crypt
data on page 0, and recovery for different pages can be executed in
undefined order.
The fix:
When delta file is created, the corresponding write filter copies only
the pages which LSN is greater then some incremental LSN. When new file
is created during incremental backup, the LSN of all it's pages must be
greater then incremental LSN, so there is no need to create delta for
such table, we can just copy it completely.
The fix is to copy the whole file which was tracked during incremental backup
with innodb ddl tracker, and copy it to base directory during --prepare
instead of delta applying.
There is also DBUG_EXECUTE_IF() in innodb code to avoid writing redo log
record for crypt data updating on page 0 to make the test case stable.
Note:
The issue is not reproducible in 10.5 as optimized DDL's are deprecated
in 10.5. But the fix is still useful because it allows to decrease
data copy size during backup, as delta file contains some extra info.
The test case should be removed for 10.5 as it will always pass.
2020-10-20 12:05:58 +02:00
|
|
|
xb_process_datadir(xtrabackup_incremental_dir, ".new",
|
|
|
|
prepare_handle_new_files, (void *)".");
|
2018-08-09 16:06:52 +02:00
|
|
|
}
|
|
|
|
else {
|
|
|
|
xb_process_datadir(".", ".new", prepare_handle_new_files);
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
int argc; for (argc = 0; argv[argc]; argc++) {}
|
2021-10-22 10:14:57 +02:00
|
|
|
xb_plugin_prepare_init(argc, argv, xtrabackup_incremental_dir);
|
2017-04-18 21:05:57 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
xtrabackup_target_dir= mysql_data_home_buff;
|
|
|
|
xtrabackup_target_dir[0]=FN_CURLIB; // all paths are relative from here
|
|
|
|
xtrabackup_target_dir[1]=0;
|
2017-09-08 00:53:21 +02:00
|
|
|
const lsn_t target_lsn = xtrabackup_incremental
|
|
|
|
? incremental_to_lsn : metadata_to_lsn;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/*
|
2017-05-26 11:51:11 +02:00
|
|
|
read metadata of target
|
2017-04-18 20:43:20 +02:00
|
|
|
*/
|
|
|
|
sprintf(metadata_path, "%s/%s", xtrabackup_target_dir,
|
|
|
|
XTRABACKUP_METADATA_FILENAME);
|
|
|
|
|
|
|
|
if (!xtrabackup_read_metadata(metadata_path)) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Error: failed to read metadata from '%s'\n",
|
2017-04-18 20:43:20 +02:00
|
|
|
metadata_path);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(false);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2017-05-26 11:51:11 +02:00
|
|
|
if (!strcmp(metadata_type, "full-backuped")) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (xtrabackup_incremental) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("error: applying incremental backup "
|
|
|
|
"needs a prepared target.");
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(false);
|
|
|
|
}
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("This target seems to be not prepared yet.");
|
2017-05-26 11:51:11 +02:00
|
|
|
} else if (!strcmp(metadata_type, "log-applied")) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("This target seems to be already prepared.");
|
2017-05-26 11:51:11 +02:00
|
|
|
} else {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("This target does not have correct metadata.");
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(false);
|
2017-05-26 11:51:11 +02:00
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
bool ok = !xtrabackup_incremental
|
|
|
|
|| metadata_to_lsn == incremental_lsn;
|
|
|
|
if (!ok) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("error: This incremental backup seems "
|
|
|
|
"not to be proper for the target. Check 'to_lsn' of the target and "
|
|
|
|
"'from_lsn' of the incremental.");
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(false);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
srv_max_n_threads = 1000;
|
2017-05-19 07:53:58 +02:00
|
|
|
srv_n_purge_threads = 1;
|
2017-04-18 21:05:57 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_filters_init();
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
srv_log_group_home_dir = NULL;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (xtrabackup_incremental) {
|
Follow-up fix to MDEV-12988 backup fails if innodb_undo_tablespaces>0
The fix broke mariabackup --prepare --incremental.
The restore of an incremental backup starts up (parts of) InnoDB twice.
First, all data files are discovered for applying .delta files. Then,
after the .delta files have been applied, InnoDB will be restarted
more completely, so that the redo log records will be applied via the
buffer pool.
During the first startup, the buffer pool is not initialized, and thus
trx_rseg_get_n_undo_tablespaces() must not be invoked. The apply of
the .delta files will currently assume that the --innodb-undo-tablespaces
option correctly specifies the number of undo tablespace files, just
like --backup does.
The second InnoDB startup of --prepare for applying the redo log will
properly invoke trx_rseg_get_n_undo_tablespaces().
enum srv_operation_mode: Add SRV_OPERATION_RESTORE_DELTA for
distinguishing the apply of .delta files from SRV_OPERATION_RESTORE.
srv_undo_tablespaces_init(): In mariabackup --prepare --incremental,
in the initial SRV_OPERATION_RESTORE_DELTA phase, do not invoke
trx_rseg_get_n_undo_tablespaces() because the buffer pool or the
redo logs are not available. Instead, blindly rely on the parameter
--innodb-undo-tablespaces.
2017-08-18 07:52:41 +02:00
|
|
|
srv_operation = SRV_OPERATION_RESTORE_DELTA;
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (innodb_init_param()) {
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
error:
|
|
|
|
ok = false;
|
|
|
|
goto cleanup;
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2019-05-24 15:19:38 +02:00
|
|
|
recv_sys.create();
|
2018-04-27 09:06:14 +02:00
|
|
|
log_sys.create();
|
2020-03-30 17:45:09 +02:00
|
|
|
recv_sys.recovery_on = true;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
xb_fil_io_init();
|
|
|
|
if (dberr_t err = xb_load_tablespaces()) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: error: xb_data_files_init() failed "
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
"with error %s\n", ut_strerr(err));
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
goto error;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
ok = fil_system.sys_space->open(false)
|
|
|
|
&& xtrabackup_apply_deltas();
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_data_files_close();
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (ok) {
|
|
|
|
/* Cleanup datadir from tablespaces deleted
|
|
|
|
between full and incremental backups */
|
|
|
|
|
|
|
|
xb_process_datadir("./", ".ibd", rm_if_not_found);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-06-18 11:26:28 +02:00
|
|
|
xb_filter_hash_free(&inc_dir_tables_hash);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
2018-03-28 08:00:06 +02:00
|
|
|
fil_system.close();
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
innodb_free_param();
|
2018-04-27 09:06:14 +02:00
|
|
|
log_sys.close();
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (!ok) goto cleanup;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2017-04-18 21:05:57 +02:00
|
|
|
|
2017-10-10 09:28:54 +02:00
|
|
|
srv_operation = xtrabackup_export
|
|
|
|
? SRV_OPERATION_RESTORE_EXPORT : SRV_OPERATION_RESTORE;
|
Follow-up fix to MDEV-12988 backup fails if innodb_undo_tablespaces>0
The fix broke mariabackup --prepare --incremental.
The restore of an incremental backup starts up (parts of) InnoDB twice.
First, all data files are discovered for applying .delta files. Then,
after the .delta files have been applied, InnoDB will be restarted
more completely, so that the redo log records will be applied via the
buffer pool.
During the first startup, the buffer pool is not initialized, and thus
trx_rseg_get_n_undo_tablespaces() must not be invoked. The apply of
the .delta files will currently assume that the --innodb-undo-tablespaces
option correctly specifies the number of undo tablespace files, just
like --backup does.
The second InnoDB startup of --prepare for applying the redo log will
properly invoke trx_rseg_get_n_undo_tablespaces().
enum srv_operation_mode: Add SRV_OPERATION_RESTORE_DELTA for
distinguishing the apply of .delta files from SRV_OPERATION_RESTORE.
srv_undo_tablespaces_init(): In mariabackup --prepare --incremental,
in the initial SRV_OPERATION_RESTORE_DELTA phase, do not invoke
trx_rseg_get_n_undo_tablespaces() because the buffer pool or the
redo logs are not available. Instead, blindly rely on the parameter
--innodb-undo-tablespaces.
2017-08-18 07:52:41 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (innodb_init_param()) {
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
goto error;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
fil_system.freeze_space_list = 0;
|
|
|
|
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Starting InnoDB instance for recovery.");
|
|
|
|
|
|
|
|
msg("mariabackup: Using %lld bytes for buffer pool "
|
|
|
|
"(set by --use-memory parameter)", xtrabackup_use_memory);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-04-18 21:05:57 +02:00
|
|
|
srv_max_buf_pool_modified_pct = (double)max_buf_pool_modified_pct;
|
|
|
|
|
|
|
|
if (srv_max_dirty_pages_pct_lwm > srv_max_buf_pool_modified_pct) {
|
|
|
|
srv_max_dirty_pages_pct_lwm = srv_max_buf_pool_modified_pct;
|
|
|
|
}
|
|
|
|
|
2020-04-02 23:43:09 +02:00
|
|
|
if (innodb_init()) {
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
goto error;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2020-10-26 14:59:30 +01:00
|
|
|
ut_ad(!fil_system.freeze_space_list);
|
|
|
|
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
corrupted_pages.read_from_file(MB_CORRUPTED_PAGES_FILE);
|
|
|
|
if (xtrabackup_incremental)
|
|
|
|
{
|
|
|
|
char inc_filename[FN_REFLEN];
|
|
|
|
sprintf(inc_filename, "%s/%s", xtrabackup_incremental_dir,
|
|
|
|
MB_CORRUPTED_PAGES_FILE);
|
|
|
|
corrupted_pages.read_from_file(inc_filename);
|
|
|
|
}
|
|
|
|
if (!corrupted_pages.empty())
|
|
|
|
corrupted_pages.zero_out_free_pages();
|
|
|
|
if (corrupted_pages.empty())
|
|
|
|
{
|
|
|
|
if (!xtrabackup_incremental && unlink(MB_CORRUPTED_PAGES_FILE) &&
|
|
|
|
errno != ENOENT)
|
|
|
|
{
|
|
|
|
char errbuf[MYSYS_STRERROR_SIZE];
|
|
|
|
my_strerror(errbuf, sizeof(errbuf), errno);
|
|
|
|
die("Error: unlink %s failed: %s", MB_CORRUPTED_PAGES_FILE,
|
|
|
|
errbuf);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
else
|
|
|
|
corrupted_pages.print_to_file(MB_CORRUPTED_PAGES_FILE);
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (ok) {
|
2019-01-17 11:33:31 +01:00
|
|
|
msg("Last binlog file %s, position %lld",
|
2018-02-20 14:10:03 +01:00
|
|
|
trx_sys.recovered_binlog_filename,
|
|
|
|
longlong(trx_sys.recovered_binlog_offset));
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Check whether the log is applied enough or not. */
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
if (recv_sys.lsn && recv_sys.lsn < target_lsn) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: error: "
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
"The log was only applied up to LSN " LSN_PF
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
", instead of " LSN_PF, recv_sys.lsn, target_lsn);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
ok = false;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2017-04-18 21:05:57 +02:00
|
|
|
#ifdef WITH_WSREP
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
else if (ok) xb_write_galera_info(xtrabackup_incremental);
|
2017-04-18 21:05:57 +02:00
|
|
|
#endif
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-04-25 20:57:52 +02:00
|
|
|
innodb_shutdown();
|
2020-04-02 23:43:09 +02:00
|
|
|
|
|
|
|
innodb_free_param();
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* output to metadata file */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (ok) {
|
2017-04-18 20:43:20 +02:00
|
|
|
char filename[FN_REFLEN];
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
strcpy(metadata_type, "log-applied");
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if(xtrabackup_incremental
|
|
|
|
&& metadata_to_lsn < incremental_to_lsn)
|
|
|
|
{
|
|
|
|
metadata_to_lsn = incremental_to_lsn;
|
|
|
|
metadata_last_lsn = incremental_last_lsn;
|
|
|
|
}
|
|
|
|
|
|
|
|
sprintf(filename, "%s/%s", xtrabackup_target_dir, XTRABACKUP_METADATA_FILENAME);
|
|
|
|
if (!xtrabackup_write_metadata(filename)) {
|
|
|
|
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: Error: failed to write metadata "
|
2019-01-14 22:28:23 +01:00
|
|
|
"to '%s'", filename);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
ok = false;
|
|
|
|
} else if (xtrabackup_extra_lsndir) {
|
2017-04-18 20:43:20 +02:00
|
|
|
sprintf(filename, "%s/%s", xtrabackup_extra_lsndir, XTRABACKUP_METADATA_FILENAME);
|
|
|
|
if (!xtrabackup_write_metadata(filename)) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: Error: failed to write "
|
2019-01-14 22:28:23 +01:00
|
|
|
"metadata to '%s'", filename);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
ok = false;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (ok) ok = apply_log_finish();
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-09-08 00:53:21 +02:00
|
|
|
if (ok && xtrabackup_export)
|
|
|
|
ok= (prepare_export() == 0);
|
|
|
|
|
MDEV-14425 Improve the redo log for concurrency
The InnoDB redo log used to be formatted in blocks of 512 bytes.
The log blocks were encrypted and the checksum was calculated while
holding log_sys.mutex, creating a serious scalability bottleneck.
We remove the fixed-size redo log block structure altogether and
essentially turn every mini-transaction into a log block of its own.
This allows encryption and checksum calculations to be performed
on local mtr_t::m_log buffers, before acquiring log_sys.mutex.
The mutex only protects a memcpy() of the data to the shared
log_sys.buf, as well as the padding of the log, in case the
to-be-written part of the log would not end in a block boundary of
the underlying storage. For now, the "padding" consists of writing
a single NUL byte, to allow recovery and mariadb-backup to detect
the end of the circular log faster.
Like the previous implementation, we will overwrite the last log block
over and over again, until it has been completely filled. It would be
possible to write only up to the last completed block (if no more
recent write was requested), or to write dummy FILE_CHECKPOINT records
to fill the incomplete block, by invoking the currently disabled
function log_pad(). This would require adjustments to some logic around
log checkpoints, page flushing, and shutdown.
An upgrade after a crash of any previous version is not supported.
Logically empty log files from a previous version will be upgraded.
An attempt to start up InnoDB without a valid ib_logfile0 will be
refused. Previously, the redo log used to be created automatically
if it was missing. Only with with innodb_force_recovery=6, it is
possible to start InnoDB in read-only mode even if the log file
does not exist. This allows the contents of a possibly corrupted
database to be dumped.
Because a prepared backup from an earlier version of mariadb-backup
will create a 0-sized log file, we will allow an upgrade from such
log files, provided that the FIL_PAGE_FILE_FLUSH_LSN in the system
tablespace looks valid.
The 512-byte log checkpoint blocks at 0x200 and 0x600 will be replaced
with 64-byte log checkpoint blocks at 0x1000 and 0x2000.
The start of log records will move from 0x800 to 0x3000. This allows us
to use 4096-byte aligned blocks for all I/O in a future revision.
We extend the MDEV-12353 redo log record format as follows.
(1) Empty mini-transactions or extra NUL bytes will not be allowed.
(2) The end-of-minitransaction marker (a NUL byte) will be replaced
with a 1-bit sequence number, which will be toggled each time when the
circular log file wraps back to the beginning.
(3) After the sequence bit, a CRC-32C checksum of all data
(excluding the sequence bit) will written.
(4) If the log is encrypted, 8 bytes will be written before
the checksum and included in it. This is part of the
initialization vector (IV) of encrypted log data.
(5) File names, page numbers, and checkpoint information will not be
encrypted. Only the payload bytes of page-level log will be encrypted.
The tablespace ID and page number will form part of the IV.
(6) For padding, arbitrary-length FILE_CHECKPOINT records may be written,
with all-zero payload, and with the normal end marker and checksum.
The minimum size is 7 bytes, or 7+8 with innodb_encrypt_log=ON.
In mariadb-backup and in Galera snapshot transfer (SST) scripts, we will
no longer remove ib_logfile0 or create an empty ib_logfile0. Server startup
will require a valid log file. When resizing the log, we will create
a logically empty ib_logfile101 at the current LSN and use an atomic rename
to replace ib_logfile0 with it. See the test innodb.log_file_size.
Because there is no mandatory padding in the log file, we are able
to create a dummy log file as of an arbitrary log sequence number.
See the test mariabackup.huge_lsn.
The parameter innodb_log_write_ahead_size and the
INFORMATION_SCHEMA.INNODB_METRICS counter log_padded will be removed.
The minimum value of innodb_log_buffer_size will be increased to 2MiB
(because log_sys.buf will replace recv_sys.buf) and the increment
adjusted to 4096 bytes (the maximum log block size).
The following INFORMATION_SCHEMA.INNODB_METRICS counters will be removed:
os_log_fsyncs
os_log_pending_fsyncs
log_pending_log_flushes
log_pending_checkpoint_writes
The following status variables will be removed:
Innodb_os_log_fsyncs (this is included in Innodb_data_fsyncs)
Innodb_os_log_pending_fsyncs (this was limited to at most 1 by design)
log_sys.get_block_size(): Return the physical block size of the log file.
This is only implemented on Linux and Microsoft Windows for now, and for
the power-of-2 block sizes between 64 and 4096 bytes (the minimum and
maximum size of a checkpoint block). If the block size is anything else,
the traditional 512-byte size will be used via normal file system
buffering.
If the file system buffers can be bypassed, a message like the following
will be issued:
InnoDB: File system buffers for log disabled (block size=512 bytes)
InnoDB: File system buffers for log disabled (block size=4096 bytes)
This has been tested on Linux and Microsoft Windows with both sizes.
On Linux, only enable O_DIRECT on the log for innodb_flush_method=O_DSYNC.
Tests in 3 different environments where the log is stored in a device
with a physical block size of 512 bytes are yielding better throughput
without O_DIRECT. This could be due to the fact that in the event the
last log block is being overwritten (if multiple transactions would
become durable at the same time, and each of will write a small
number of bytes to the last log block), it should be faster to re-copy
data from log_sys.buf or log_sys.flush_buf to the kernel buffer,
to be finally written at fdatasync() time.
The parameter innodb_flush_method=O_DSYNC will imply O_DIRECT for
data files. This option will enable O_DIRECT on the log file on Linux.
It may be unsafe to use when the storage device does not support
FUA (Force Unit Access) mode.
When the server is compiled WITH_PMEM=ON, we will use memory-mapped
I/O for the log file if the log resides on a "mount -o dax" device.
We will identify PMEM in a start-up message:
InnoDB: log sequence number 0 (memory-mapped); transaction id 3
On Linux, we will also invoke mmap() on any ib_logfile0 that resides
in /dev/shm, effectively treating the log file as persistent memory.
This should speed up "./mtr --mem" and increase the test coverage of
PMEM on non-PMEM hardware. It also allows users to estimate how much
the performance would be improved by installing persistent memory.
On other tmpfs file systems such as /run, we will not use mmap().
mariadb-backup: Eliminated several variables. We will refer
directly to recv_sys and log_sys.
backup_wait_for_lsn(): Detect non-progress of
xtrabackup_copy_logfile(). In this new log format with
arbitrary-sized blocks, we can only detect log file overrun
indirectly, by observing that the scanned log sequence number
is not advancing.
xtrabackup_copy_logfile(): On PMEM, do not modify the sequence bit,
because we are not allowed to modify the server's log file, and our
memory mapping is read-only.
trx_flush_log_if_needed_low(): Do not use the callback on pmem.
Using neither flush_lock nor write_lock around PMEM writes seems
to yield the best performance. The pmem_persist() calls may
still be somewhat slower than the pwrite() and fdatasync() based
interface (PMEM mounted without -o dax).
recv_sys_t::buf: Remove. We will use log_sys.buf for parsing.
recv_sys_t::MTR_SIZE_MAX: Replaces RECV_SCAN_SIZE.
recv_sys_t::file_checkpoint: Renamed from mlog_checkpoint_lsn.
recv_sys_t, log_sys_t: Removed many data members.
recv_sys.lsn: Renamed from recv_sys.recovered_lsn.
recv_sys.offset: Renamed from recv_sys.recovered_offset.
log_sys.buf_size: Replaces srv_log_buffer_size.
recv_buf: A smart pointer that wraps log_sys.buf[recv_sys.offset]
when the buffer is being allocated from the memory heap.
recv_ring: A smart pointer that wraps a circular log_sys.buf[] that is
backed by ib_logfile0. The pointer will wrap from recv_sys.len
(log_sys.file_size) to log_sys.START_OFFSET. For the record that
wraps around, we may copy file name or record payload data to
the auxiliary buffer decrypt_buf in order to have a contiguous
block of memory. The maximum size of a record is less than
innodb_page_size bytes.
recv_sys_t::parse(): Take the smart pointer as a template parameter.
Do not temporarily add a trailing NUL byte to FILE_ records, because
we are not supposed to modify the memory-mapped log file. (It is
attached in read-write mode already during recovery.)
recv_sys_t::parse_mtr(): Wrapper for recv_sys_t::parse().
recv_sys_t::parse_pmem(): Like parse_mtr(), but if PREMATURE_EOF would be
returned on PMEM, use recv_ring to wrap around the buffer to the start.
mtr_t::finish_write(), log_close(): Do not enforce log_sys.max_buf_free
on PMEM, because it has no meaning on the mmap-based log.
log_sys.write_to_buf: Count writes to log_sys.buf. Replaces
srv_stats.log_write_requests and export_vars.innodb_log_write_requests.
Protected by log_sys.mutex. Updated consistently in log_close().
Previously, mtr_t::commit() conditionally updated the count,
which was inconsistent.
log_sys.write_to_log: Count swaps of log_sys.buf and log_sys.flush_buf,
for writing to log_sys.log (the ib_logfile0). Replaces
srv_stats.log_writes and export_vars.innodb_log_writes.
Protected by log_sys.mutex.
log_sys.waits: Count waits in append_prepare(). Replaces
srv_stats.log_waits and export_vars.innodb_log_waits.
recv_recover_page(): Do not unnecessarily acquire
log_sys.flush_order_mutex. We are inserting the blocks in arbitary
order anyway, to be adjusted in recv_sys.apply(true).
We will change the definition of flush_lock and write_lock to
avoid potential false sharing. Depending on sizeof(log_sys) and
CPU_LEVEL1_DCACHE_LINESIZE, the flush_lock and write_lock could
share a cache line with each other or with the last data members
of log_sys.
Thanks to Matthias Leich for providing https://rr-project.org traces
for various failures during the development, and to
Thirunarayanan Balathandayuthapani for his help in debugging
some of the recovery code. And thanks to the developers of the
rr debugger for a tool without which extensive changes to InnoDB
would be very challenging to get right.
Thanks to Vladislav Vaintroub for useful feedback and
to him, Axel Schwenke and Krunal Bauskar for testing the performance.
2022-01-21 15:03:47 +01:00
|
|
|
cleanup:
|
2017-04-18 20:43:20 +02:00
|
|
|
xb_filters_free();
|
MDEV-22929 MariaBackup option to report and/or continue when corruption is encountered
The new option --log-innodb-page-corruption is introduced.
When this option is set, backup is not interrupted if innodb corrupted
page is detected. Instead it logs all found corrupted pages in
innodb_corrupted_pages file in backup directory and finishes with error.
For incremental backup corrupted pages are also copied to .delta file,
because we can't do LSN check for such pages during backup,
innodb_corrupted_pages will also be created in incremental backup
directory.
During --prepare, corrupted pages list is read from the file just after
redo log is applied, and each page from the list is checked if it is allocated
in it's tablespace or not. If it is not allocated, then it is zeroed out,
flushed to the tablespace and removed from the list. If all pages are removed
from the list, then --prepare is finished successfully and
innodb_corrupted_pages file is removed from backup directory. Otherwise
--prepare is finished with error message and innodb_corrupted_pages contains
the list of the pages, which are detected as corrupted during backup, and are
allocated in their tablespaces, what means backup directory contains corrupted
innodb pages, and backup can not be considered as consistent.
For incremental --prepare corrupted pages from .delta files are applied
to the base backup, innodb_corrupted_pages is read from both base in
incremental directories, and the same action is proceded for corrupted
pages list as for full --prepare. innodb_corrupted_pages file is
modified or removed only in base directory.
If DDL happens during backup, it is also processed at the end of backup
to have correct tablespace names in innodb_corrupted_pages.
2020-08-20 15:49:40 +02:00
|
|
|
return ok && !ib::error::was_logged() && corrupted_pages.empty();
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/**************************************************************************
|
|
|
|
Append group name to xb_load_default_groups list. */
|
|
|
|
static
|
|
|
|
void
|
|
|
|
append_defaults_group(const char *group, const char *default_groups[],
|
|
|
|
size_t default_groups_size)
|
|
|
|
{
|
|
|
|
uint i;
|
|
|
|
bool appended = false;
|
|
|
|
for (i = 0; i < default_groups_size - 1; i++) {
|
|
|
|
if (default_groups[i] == NULL) {
|
|
|
|
default_groups[i] = group;
|
|
|
|
appended = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
ut_a(appended);
|
|
|
|
}
|
|
|
|
|
2018-12-14 23:36:21 +01:00
|
|
|
static const char*
|
|
|
|
normalize_privilege_target_name(const char* name)
|
|
|
|
{
|
|
|
|
if (strcmp(name, "*") == 0) {
|
|
|
|
return "\\*";
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
/* should have no regex special characters. */
|
|
|
|
ut_ad(strpbrk(name, ".()[]*+?") == 0);
|
|
|
|
}
|
|
|
|
return name;
|
|
|
|
}
|
|
|
|
|
|
|
|
/******************************************************************//**
|
|
|
|
Check if specific privilege is granted.
|
|
|
|
Uses regexp magic to check if requested privilege is granted for given
|
|
|
|
database.table or database.* or *.*
|
|
|
|
or if user has 'ALL PRIVILEGES' granted.
|
|
|
|
@return true if requested privilege is granted, false otherwise. */
|
|
|
|
static bool
|
|
|
|
has_privilege(const std::list<std::string> &granted,
|
|
|
|
const char* required,
|
|
|
|
const char* db_name,
|
|
|
|
const char* table_name)
|
|
|
|
{
|
|
|
|
char buffer[1000];
|
|
|
|
regex_t priv_re;
|
|
|
|
regmatch_t tables_regmatch[1];
|
|
|
|
bool result = false;
|
|
|
|
|
|
|
|
db_name = normalize_privilege_target_name(db_name);
|
|
|
|
table_name = normalize_privilege_target_name(table_name);
|
|
|
|
|
|
|
|
int written = snprintf(buffer, sizeof(buffer),
|
|
|
|
"GRANT .*(%s)|(ALL PRIVILEGES).* ON (\\*|`%s`)\\.(\\*|`%s`)",
|
|
|
|
required, db_name, table_name);
|
|
|
|
if (written < 0 || written == sizeof(buffer)
|
|
|
|
|| regcomp(&priv_re, buffer, REG_EXTENDED)) {
|
2019-01-15 22:47:54 +01:00
|
|
|
die("regcomp() failed for '%s'", buffer);
|
2018-12-14 23:36:21 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
typedef std::list<std::string>::const_iterator string_iter;
|
|
|
|
for (string_iter i = granted.begin(), e = granted.end(); i != e; ++i) {
|
|
|
|
int res = regexec(&priv_re, i->c_str(),
|
|
|
|
1, tables_regmatch, 0);
|
|
|
|
|
|
|
|
if (res != REG_NOMATCH) {
|
|
|
|
result = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
xb_regfree(&priv_re);
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
enum {
|
|
|
|
PRIVILEGE_OK = 0,
|
|
|
|
PRIVILEGE_WARNING = 1,
|
|
|
|
PRIVILEGE_ERROR = 2,
|
|
|
|
};
|
|
|
|
|
|
|
|
/******************************************************************//**
|
|
|
|
Check if specific privilege is granted.
|
|
|
|
Prints error message if required privilege is missing.
|
|
|
|
@return PRIVILEGE_OK if requested privilege is granted, error otherwise. */
|
|
|
|
static
|
|
|
|
int check_privilege(
|
|
|
|
const std::list<std::string> &granted_priv, /* in: list of
|
|
|
|
granted privileges*/
|
|
|
|
const char* required, /* in: required privilege name */
|
|
|
|
const char* target_database, /* in: required privilege target
|
|
|
|
database name */
|
|
|
|
const char* target_table, /* in: required privilege target
|
|
|
|
table name */
|
|
|
|
int error = PRIVILEGE_ERROR) /* in: return value if privilege
|
|
|
|
is not granted */
|
|
|
|
{
|
|
|
|
if (!has_privilege(granted_priv,
|
|
|
|
required, target_database, target_table)) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("%s: missing required privilege %s on %s.%s",
|
2018-12-14 23:36:21 +01:00
|
|
|
(error == PRIVILEGE_ERROR ? "Error" : "Warning"),
|
|
|
|
required, target_database, target_table);
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
return PRIVILEGE_OK;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2018-12-28 14:33:34 +01:00
|
|
|
/**
|
2018-12-14 23:36:21 +01:00
|
|
|
Check DB user privileges according to the intended actions.
|
|
|
|
|
|
|
|
Fetches DB user privileges, determines intended actions based on
|
|
|
|
command-line arguments and prints missing privileges.
|
2018-12-28 14:33:34 +01:00
|
|
|
@return whether all the necessary privileges are granted */
|
|
|
|
static bool check_all_privileges()
|
2018-12-14 23:36:21 +01:00
|
|
|
{
|
|
|
|
if (!mysql_connection) {
|
|
|
|
/* Not connected, no queries is going to be executed. */
|
2018-12-28 14:33:34 +01:00
|
|
|
return true;
|
2018-12-14 23:36:21 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Fetch effective privileges. */
|
|
|
|
std::list<std::string> granted_privileges;
|
|
|
|
MYSQL_RES* result = xb_mysql_query(mysql_connection, "SHOW GRANTS",
|
2018-12-28 14:33:34 +01:00
|
|
|
true);
|
|
|
|
while (MYSQL_ROW row = mysql_fetch_row(result)) {
|
2018-12-14 23:36:21 +01:00
|
|
|
granted_privileges.push_back(*row);
|
|
|
|
}
|
|
|
|
mysql_free_result(result);
|
|
|
|
|
|
|
|
int check_result = PRIVILEGE_OK;
|
|
|
|
|
|
|
|
/* FLUSH TABLES WITH READ LOCK */
|
|
|
|
if (!opt_no_lock)
|
|
|
|
{
|
|
|
|
check_result |= check_privilege(
|
|
|
|
granted_privileges,
|
|
|
|
"RELOAD", "*", "*");
|
|
|
|
check_result |= check_privilege(
|
|
|
|
granted_privileges,
|
2018-12-28 14:33:34 +01:00
|
|
|
"PROCESS", "*", "*");
|
2018-12-14 23:36:21 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/* KILL ... */
|
2022-07-29 16:11:08 +02:00
|
|
|
if (!opt_no_lock && (opt_kill_long_queries_timeout || opt_kill_long_query_type)) {
|
2018-12-14 23:36:21 +01:00
|
|
|
check_result |= check_privilege(
|
|
|
|
granted_privileges,
|
2022-07-29 16:11:08 +02:00
|
|
|
"CONNECTION ADMIN", "*", "*",
|
|
|
|
PRIVILEGE_WARNING);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* START SLAVE SQL_THREAD */
|
|
|
|
/* STOP SLAVE SQL_THREAD */
|
|
|
|
if (opt_safe_slave_backup) {
|
|
|
|
check_result |= check_privilege(
|
|
|
|
granted_privileges,
|
|
|
|
"REPLICATION SLAVE ADMIN", "*", "*",
|
2018-12-14 23:36:21 +01:00
|
|
|
PRIVILEGE_WARNING);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* SHOW MASTER STATUS */
|
|
|
|
/* SHOW SLAVE STATUS */
|
|
|
|
if (opt_galera_info || opt_slave_info
|
2022-07-29 16:11:08 +02:00
|
|
|
|| opt_safe_slave_backup) {
|
2018-12-14 23:36:21 +01:00
|
|
|
check_result |= check_privilege(granted_privileges,
|
2022-07-29 16:11:08 +02:00
|
|
|
"SLAVE MONITOR", "*", "*",
|
2018-12-14 23:36:21 +01:00
|
|
|
PRIVILEGE_WARNING);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (check_result & PRIVILEGE_ERROR) {
|
2019-05-02 15:25:24 +02:00
|
|
|
msg("Current privileges, as reported by 'SHOW GRANTS': ");
|
|
|
|
int n=1;
|
|
|
|
for (std::list<std::string>::const_iterator it = granted_privileges.begin();
|
|
|
|
it != granted_privileges.end();
|
|
|
|
it++,n++) {
|
|
|
|
msg(" %d.%s", n, it->c_str());
|
|
|
|
}
|
2019-05-05 14:03:48 +02:00
|
|
|
return false;
|
2018-12-14 23:36:21 +01:00
|
|
|
}
|
2019-05-05 14:03:48 +02:00
|
|
|
|
|
|
|
return true;
|
2018-12-14 23:36:21 +01:00
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
bool
|
|
|
|
xb_init()
|
|
|
|
{
|
|
|
|
const char *mixed_options[4] = {NULL, NULL, NULL, NULL};
|
|
|
|
int n_mixed_options;
|
|
|
|
|
|
|
|
/* sanity checks */
|
|
|
|
|
|
|
|
if (opt_slave_info
|
|
|
|
&& opt_no_lock
|
|
|
|
&& !opt_safe_slave_backup) {
|
|
|
|
msg("Error: --slave-info is used with --no-lock but "
|
|
|
|
"without --safe-slave-backup. The binlog position "
|
2019-01-14 22:28:23 +01:00
|
|
|
"cannot be consistent with the backup data.");
|
2017-04-18 20:43:20 +02:00
|
|
|
return(false);
|
|
|
|
}
|
|
|
|
|
2018-01-22 21:01:03 +01:00
|
|
|
if (xtrabackup_backup && opt_rsync)
|
|
|
|
{
|
|
|
|
if (xtrabackup_stream_fmt)
|
|
|
|
{
|
|
|
|
msg("Error: --rsync doesn't work with --stream\n");
|
|
|
|
return(false);
|
|
|
|
}
|
|
|
|
bool have_rsync = IF_WIN(false, (system("rsync --version > /dev/null 2>&1") == 0));
|
|
|
|
if (!have_rsync)
|
|
|
|
{
|
|
|
|
msg("Error: rsync executable not found, cannot run backup with --rsync\n");
|
|
|
|
return false;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
n_mixed_options = 0;
|
|
|
|
|
|
|
|
if (opt_decompress) {
|
|
|
|
mixed_options[n_mixed_options++] = "--decompress";
|
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_copy_back) {
|
|
|
|
mixed_options[n_mixed_options++] = "--copy-back";
|
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_move_back) {
|
|
|
|
mixed_options[n_mixed_options++] = "--move-back";
|
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_prepare) {
|
|
|
|
mixed_options[n_mixed_options++] = "--apply-log";
|
|
|
|
}
|
|
|
|
|
|
|
|
if (n_mixed_options > 1) {
|
|
|
|
msg("Error: %s and %s are mutually exclusive\n",
|
|
|
|
mixed_options[0], mixed_options[1]);
|
|
|
|
return(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_backup) {
|
|
|
|
if ((mysql_connection = xb_mysql_connect()) == NULL) {
|
|
|
|
return(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!get_mysql_vars(mysql_connection)) {
|
|
|
|
return(false);
|
|
|
|
}
|
2022-02-25 12:00:48 +01:00
|
|
|
|
2018-12-28 14:33:34 +01:00
|
|
|
if (opt_check_privileges && !check_all_privileges()) {
|
|
|
|
return(false);
|
2018-12-14 23:36:21 +01:00
|
|
|
}
|
2022-02-25 12:00:48 +01:00
|
|
|
history_start_time = time(NULL);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
return(true);
|
|
|
|
}
|
|
|
|
|
2017-04-18 21:05:57 +02:00
|
|
|
|
|
|
|
extern void init_signals(void);
|
|
|
|
|
|
|
|
#include <sql_locale.h>
|
|
|
|
|
2019-12-11 11:22:03 +01:00
|
|
|
|
2017-04-18 21:05:57 +02:00
|
|
|
void setup_error_messages()
|
|
|
|
{
|
|
|
|
my_default_lc_messages = &my_locale_en_US;
|
2019-12-11 11:22:03 +01:00
|
|
|
if (init_errmessage())
|
|
|
|
die("could not initialize error messages");
|
2017-04-18 21:05:57 +02:00
|
|
|
}
|
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
/** Handle mariabackup options. The options are handled with the following
|
|
|
|
order:
|
|
|
|
|
|
|
|
1) Load server groups and process server options, ignore unknown options
|
|
|
|
2) Load client groups and process client options, ignore unknown options
|
|
|
|
3) Load backup groups and process client-server options, exit on unknown option
|
|
|
|
4) Process --mysqld-args options, ignore unknown options
|
|
|
|
|
|
|
|
@param[in] argc arguments count
|
|
|
|
@param[in] argv arguments array
|
|
|
|
@param[out] argv_server server options including loaded from server groups
|
|
|
|
@param[out] argv_client client options including loaded from client groups
|
|
|
|
@param[out] argv_backup backup options including loaded from backup groups */
|
|
|
|
void handle_options(int argc, char **argv, char ***argv_server,
|
|
|
|
char ***argv_client, char ***argv_backup)
|
2017-04-18 20:43:20 +02:00
|
|
|
{
|
2017-04-18 21:05:57 +02:00
|
|
|
/* Setup some variables for Innodb.*/
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
srv_operation = SRV_OPERATION_RESTORE;
|
2017-04-18 21:05:57 +02:00
|
|
|
|
2019-06-28 07:05:12 +02:00
|
|
|
files_charset_info = &my_charset_utf8mb3_general_ci;
|
2018-08-09 16:06:52 +02:00
|
|
|
|
2017-04-18 21:05:57 +02:00
|
|
|
|
|
|
|
setup_error_messages();
|
|
|
|
sys_var_init();
|
|
|
|
plugin_mutex_init();
|
2018-01-27 18:46:31 +01:00
|
|
|
mysql_prlock_init(key_rwlock_LOCK_system_variables_hash, &LOCK_system_variables_hash);
|
2017-04-18 21:05:57 +02:00
|
|
|
opt_stack_trace = 1;
|
|
|
|
test_flags |= TEST_SIGINT;
|
|
|
|
init_signals();
|
|
|
|
#ifndef _WIN32
|
|
|
|
/* Exit process on SIGINT. */
|
|
|
|
my_sigset(SIGINT, SIG_DFL);
|
|
|
|
#endif
|
|
|
|
|
2017-09-08 08:33:41 +02:00
|
|
|
sf_leaking_memory = 1; /* don't report memory leaks on early exist */
|
2017-04-18 21:05:57 +02:00
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
int i;
|
|
|
|
int ho_error;
|
|
|
|
|
|
|
|
char* target_dir = NULL;
|
|
|
|
bool prepare = false;
|
|
|
|
|
|
|
|
char conf_file[FN_REFLEN];
|
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
// array_elements() will not work for load_defaults, as it is defined
|
|
|
|
// as external symbol, so let's use dynamic array to have ability to
|
|
|
|
// add new server default groups
|
|
|
|
std::vector<const char *> server_default_groups;
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
for (const char **default_group= load_default_groups; *default_group;
|
|
|
|
++default_group)
|
|
|
|
server_default_groups.push_back(*default_group);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
std::vector<char *> mysqld_args;
|
|
|
|
std::vector<char *> mariabackup_args;
|
|
|
|
mysqld_args.push_back(argv[0]);
|
|
|
|
mariabackup_args.push_back(argv[0]);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
/* scan options for group and config file to load defaults from */
|
|
|
|
for (i= 1; i < argc; i++)
|
|
|
|
{
|
|
|
|
char *optend= strcend(argv[i], '=');
|
|
|
|
if (mysqld_args.size() > 1 ||
|
|
|
|
strncmp(argv[i], "--mysqld-args", optend - argv[i]) == 0)
|
|
|
|
{
|
|
|
|
mysqld_args.push_back(argv[i]);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
mariabackup_args.push_back(argv[i]);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
if (strncmp(argv[i], "--defaults-group", optend - argv[i]) == 0)
|
|
|
|
{
|
|
|
|
defaults_group= optend + 1;
|
|
|
|
server_default_groups.push_back(defaults_group);
|
|
|
|
}
|
|
|
|
else if (strncmp(argv[i], "--login-path", optend - argv[i]) == 0)
|
|
|
|
{
|
|
|
|
append_defaults_group(optend + 1, xb_client_default_groups,
|
|
|
|
array_elements(xb_client_default_groups));
|
|
|
|
}
|
|
|
|
else if (!strncmp(argv[i], "--prepare", optend - argv[i]))
|
|
|
|
{
|
|
|
|
prepare= true;
|
|
|
|
}
|
|
|
|
else if (!strncmp(argv[i], "--apply-log", optend - argv[i]))
|
|
|
|
{
|
|
|
|
prepare= true;
|
|
|
|
}
|
2021-10-11 14:53:11 +02:00
|
|
|
else if (!strncmp(argv[i], "--incremental-dir", optend - argv[i]) &&
|
2020-06-08 10:45:56 +02:00
|
|
|
*optend)
|
|
|
|
{
|
|
|
|
target_dir= optend + 1;
|
|
|
|
}
|
2021-10-11 14:53:11 +02:00
|
|
|
else if (!strncmp(argv[i], "--target-dir", optend - argv[i]) &&
|
|
|
|
*optend && !target_dir)
|
|
|
|
{
|
|
|
|
target_dir= optend + 1;
|
|
|
|
}
|
|
|
|
else if (!*optend && argv[i][0] != '-' && !target_dir)
|
2020-06-08 10:45:56 +02:00
|
|
|
{
|
|
|
|
target_dir= argv[i];
|
|
|
|
}
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
server_default_groups.push_back(NULL);
|
2017-04-18 20:43:20 +02:00
|
|
|
snprintf(conf_file, sizeof(conf_file), "my");
|
|
|
|
|
|
|
|
if (prepare && target_dir) {
|
|
|
|
snprintf(conf_file, sizeof(conf_file),
|
|
|
|
"%s/backup-my.cnf", target_dir);
|
2017-04-18 21:05:57 +02:00
|
|
|
if (!strncmp(argv[1], "--defaults-file=", 16)) {
|
|
|
|
/* Remove defaults-file*/
|
|
|
|
for (int i = 2; ; i++) {
|
|
|
|
if ((argv[i-1]= argv[i]) == 0)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
argc--;
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2017-04-18 21:05:57 +02:00
|
|
|
|
2021-03-08 19:54:58 +01:00
|
|
|
mariabackup_args.push_back(nullptr);
|
2020-06-08 10:45:56 +02:00
|
|
|
*argv_client= *argv_server= *argv_backup= &mariabackup_args[0];
|
2021-03-08 19:54:58 +01:00
|
|
|
int argc_backup= static_cast<int>(mariabackup_args.size() - 1);
|
2020-06-08 10:45:56 +02:00
|
|
|
int argc_client= argc_backup;
|
|
|
|
int argc_server= argc_backup;
|
|
|
|
|
|
|
|
/* 1) Load server groups and process server options, ignore unknown
|
|
|
|
options */
|
|
|
|
|
|
|
|
load_defaults_or_exit(conf_file, &server_default_groups[0],
|
|
|
|
&argc_server, argv_server);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-04-18 21:05:57 +02:00
|
|
|
int n;
|
|
|
|
for (n = 0; (*argv_server)[n]; n++) {};
|
|
|
|
argc_server = n;
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
print_param_str <<
|
|
|
|
"# This MySQL options file was generated by XtraBackup.\n"
|
|
|
|
"[" << defaults_group << "]\n";
|
|
|
|
|
|
|
|
/* We want xtrabackup to ignore unknown options, because it only
|
|
|
|
recognizes a small subset of server variables */
|
|
|
|
my_getopt_skip_unknown = TRUE;
|
|
|
|
|
|
|
|
/* Reset u_max_value for all options, as we don't want the
|
|
|
|
--maximum-... modifier to set the actual option values */
|
|
|
|
for (my_option *optp= xb_server_options; optp->name; optp++) {
|
|
|
|
optp->u_max_value = (G_PTR *) &global_max_value;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Throw a descriptive error if --defaults-file or --defaults-extra-file
|
|
|
|
is not the first command line argument */
|
|
|
|
for (int i = 2 ; i < argc ; i++) {
|
|
|
|
char *optend = strcend((argv)[i], '=');
|
|
|
|
|
|
|
|
if (optend - argv[i] == 15 &&
|
2019-01-15 22:47:54 +01:00
|
|
|
!strncmp(argv[i], "--defaults-file", optend - argv[i])) {
|
|
|
|
die("--defaults-file must be specified first on the command line");
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2019-01-15 22:47:54 +01:00
|
|
|
if (optend - argv[i] == 21 &&
|
|
|
|
!strncmp(argv[i], "--defaults-extra-file",
|
|
|
|
optend - argv[i])) {
|
|
|
|
die("--defaults-extra-file must be specified first on the command line");
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (argc_server > 0
|
|
|
|
&& (ho_error=handle_options(&argc_server, argv_server,
|
|
|
|
xb_server_options, xb_get_one_option)))
|
|
|
|
exit(ho_error);
|
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
/* 2) Load client groups and process client options, ignore unknown
|
|
|
|
options */
|
|
|
|
|
2018-03-16 13:35:42 +01:00
|
|
|
load_defaults_or_exit(conf_file, xb_client_default_groups,
|
|
|
|
&argc_client, argv_client);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-04-18 21:05:57 +02:00
|
|
|
for (n = 0; (*argv_client)[n]; n++) {};
|
|
|
|
argc_client = n;
|
|
|
|
|
2017-07-13 16:49:57 +02:00
|
|
|
if (innobackupex_mode && argc_client > 0) {
|
2017-04-18 20:43:20 +02:00
|
|
|
if (!ibx_handle_options(&argc_client, argv_client)) {
|
|
|
|
exit(EXIT_FAILURE);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (argc_client > 0
|
|
|
|
&& (ho_error=handle_options(&argc_client, argv_client,
|
|
|
|
xb_client_options, xb_get_one_option)))
|
|
|
|
exit(ho_error);
|
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
/* 3) Load backup groups and process client-server options, exit on
|
|
|
|
unknown option */
|
|
|
|
|
|
|
|
load_defaults_or_exit(conf_file, backup_default_groups, &argc_backup,
|
|
|
|
argv_backup);
|
|
|
|
for (n= 0; (*argv_backup)[n]; n++)
|
|
|
|
{
|
|
|
|
};
|
|
|
|
argc_backup= n;
|
|
|
|
|
|
|
|
my_handle_options_init_variables = FALSE;
|
|
|
|
|
|
|
|
if (argc_backup > 0 &&
|
|
|
|
(ho_error= handle_options(&argc_backup, argv_backup,
|
|
|
|
xb_server_options, xb_get_one_option)))
|
|
|
|
exit(ho_error);
|
|
|
|
|
|
|
|
/* Add back the program name handle_options removes */
|
|
|
|
++argc_backup;
|
|
|
|
--(*argv_backup);
|
|
|
|
|
|
|
|
if (innobackupex_mode && argc_backup > 0 &&
|
|
|
|
!ibx_handle_options(&argc_backup, argv_backup))
|
|
|
|
exit(EXIT_FAILURE);
|
|
|
|
|
|
|
|
my_getopt_skip_unknown = FALSE;
|
|
|
|
|
|
|
|
if (argc_backup > 0 &&
|
|
|
|
(ho_error= handle_options(&argc_backup, argv_backup,
|
|
|
|
xb_client_options, xb_get_one_option)))
|
|
|
|
exit(ho_error);
|
|
|
|
|
|
|
|
if (opt_password)
|
|
|
|
{
|
2021-02-05 13:57:46 +01:00
|
|
|
char *argument= (char*) opt_password;
|
|
|
|
char *start= (char*) opt_password;
|
|
|
|
opt_password= my_strdup(PSI_NOT_INSTRUMENTED, opt_password,
|
|
|
|
MYF(MY_FAE));
|
2020-06-08 10:45:56 +02:00
|
|
|
while (*argument)
|
|
|
|
*argument++= 'x'; // Destroy argument
|
|
|
|
if (*start)
|
|
|
|
start[1]= 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* 4) Process --mysqld-args options, ignore unknown options */
|
|
|
|
|
|
|
|
my_getopt_skip_unknown = TRUE;
|
|
|
|
|
|
|
|
int argc_mysqld = static_cast<int>(mysqld_args.size());
|
|
|
|
if (argc_mysqld > 1)
|
|
|
|
{
|
|
|
|
char **argv_mysqld= &mysqld_args[0];
|
|
|
|
if ((ho_error= handle_options(&argc_mysqld, &argv_mysqld,
|
|
|
|
xb_server_options, xb_get_one_option)))
|
|
|
|
exit(ho_error);
|
|
|
|
}
|
|
|
|
|
|
|
|
my_handle_options_init_variables = TRUE;
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* Reject command line arguments that don't look like options, i.e. are
|
|
|
|
not of the form '-X' (single-character options) or '--option' (long
|
|
|
|
options) */
|
2020-06-08 10:45:56 +02:00
|
|
|
for (int i = 0 ; i < argc_backup ; i++) {
|
|
|
|
const char * const opt = (*argv_backup)[i];
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (strncmp(opt, "--", 2) &&
|
|
|
|
!(strlen(opt) == 2 && opt[0] == '-')) {
|
|
|
|
bool server_option = true;
|
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
for (int j = 0; j < argc_backup; j++) {
|
|
|
|
if (opt == (*argv_backup)[j]) {
|
2017-04-18 20:43:20 +02:00
|
|
|
server_option = false;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!server_option) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: Error:"
|
2019-01-14 22:28:23 +01:00
|
|
|
" unknown argument: '%s'", opt);
|
2017-04-18 20:43:20 +02:00
|
|
|
exit(EXIT_FAILURE);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
static int main_low(char** argv);
|
2017-09-08 00:53:21 +02:00
|
|
|
static int get_exepath(char *buf, size_t size, const char *argv0);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
/* ================= main =================== */
|
2017-04-18 20:43:20 +02:00
|
|
|
int main(int argc, char **argv)
|
|
|
|
{
|
2020-06-08 10:45:56 +02:00
|
|
|
char **server_defaults;
|
|
|
|
char **client_defaults;
|
|
|
|
char **backup_defaults;
|
2017-09-08 00:53:21 +02:00
|
|
|
|
2021-04-07 16:44:30 +02:00
|
|
|
my_getopt_prefix_matching= 0;
|
|
|
|
|
2017-09-08 00:53:21 +02:00
|
|
|
if (get_exepath(mariabackup_exe,FN_REFLEN, argv[0]))
|
|
|
|
strncpy(mariabackup_exe,argv[0], FN_REFLEN-1);
|
|
|
|
|
|
|
|
|
|
|
|
if (argc > 1 )
|
2017-04-18 21:05:57 +02:00
|
|
|
{
|
2017-09-08 00:53:21 +02:00
|
|
|
/* In "prepare export", we need to start mysqld
|
|
|
|
Since it is not always be installed on the machine,
|
|
|
|
we start "mariabackup --mysqld", which acts as mysqld
|
|
|
|
*/
|
|
|
|
if (strcmp(argv[1], "--mysqld") == 0)
|
|
|
|
{
|
|
|
|
extern int mysqld_main(int argc, char **argv);
|
|
|
|
argc--;
|
|
|
|
argv++;
|
|
|
|
argv[0]+=2;
|
|
|
|
return mysqld_main(argc, argv);
|
|
|
|
}
|
|
|
|
if(strcmp(argv[1], "--innobackupex") == 0)
|
|
|
|
{
|
|
|
|
argv++;
|
|
|
|
argc--;
|
|
|
|
innobackupex_mode = true;
|
|
|
|
}
|
2017-04-18 21:05:57 +02:00
|
|
|
}
|
2017-09-08 00:53:21 +02:00
|
|
|
|
|
|
|
if (argc > 1)
|
|
|
|
strncpy(orig_argv1,argv[1],sizeof(orig_argv1) -1);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
2017-04-18 21:05:57 +02:00
|
|
|
init_signals();
|
2017-04-18 20:43:20 +02:00
|
|
|
MY_INIT(argv[0]);
|
|
|
|
|
|
|
|
xb_regex_init();
|
|
|
|
|
|
|
|
capture_tool_command(argc, argv);
|
|
|
|
|
|
|
|
if (mysql_server_init(-1, NULL, NULL))
|
|
|
|
{
|
2019-01-15 22:47:54 +01:00
|
|
|
die("mysql_server_init() failed");
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
2019-06-28 07:05:12 +02:00
|
|
|
system_charset_info = &my_charset_utf8mb3_general_ci;
|
2017-04-18 20:43:20 +02:00
|
|
|
key_map_full.set_all();
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
logger.init_base();
|
2022-04-12 13:07:20 +02:00
|
|
|
logger.set_handlers(LOG_NONE, LOG_NONE);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
mysql_mutex_init(key_LOCK_error_log, &LOCK_error_log,
|
|
|
|
MY_MUTEX_INIT_FAST);
|
|
|
|
|
2020-06-08 10:45:56 +02:00
|
|
|
handle_options(argc, argv, &server_defaults, &client_defaults,
|
|
|
|
&backup_defaults);
|
2017-04-18 20:43:20 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
#ifndef DBUG_OFF
|
|
|
|
if (dbug_option) {
|
|
|
|
DBUG_SET_INITIAL(dbug_option);
|
|
|
|
DBUG_SET(dbug_option);
|
|
|
|
}
|
|
|
|
#endif
|
2019-10-29 22:37:12 +01:00
|
|
|
/* Main functions for library */
|
|
|
|
init_thr_timer(5);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
|
|
|
int status = main_low(server_defaults);
|
|
|
|
|
2019-10-29 22:37:12 +01:00
|
|
|
end_thr_timer();
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
backup_cleanup();
|
|
|
|
|
|
|
|
if (innobackupex_mode) {
|
|
|
|
ibx_cleanup();
|
|
|
|
}
|
|
|
|
|
|
|
|
free_defaults(server_defaults);
|
2020-06-08 10:45:56 +02:00
|
|
|
free_defaults(client_defaults);
|
|
|
|
free_defaults(backup_defaults);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
|
|
|
|
#ifndef DBUG_OFF
|
|
|
|
if (dbug_option) {
|
|
|
|
DBUG_END();
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
logger.cleanup_base();
|
2019-12-11 11:22:03 +01:00
|
|
|
cleanup_errmsgs();
|
|
|
|
free_error_messages();
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
mysql_mutex_destroy(&LOCK_error_log);
|
2017-04-18 21:05:57 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (status == EXIT_SUCCESS) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("completed OK!");
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
}
|
2017-04-18 21:05:57 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return status;
|
|
|
|
}
|
2017-04-18 21:05:57 +02:00
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
static int main_low(char** argv)
|
|
|
|
{
|
2017-04-18 20:43:20 +02:00
|
|
|
if (innobackupex_mode) {
|
|
|
|
if (!ibx_init()) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (!xtrabackup_print_param && !xtrabackup_prepare
|
|
|
|
&& !strcmp(mysql_data_home, "./")) {
|
2017-04-18 20:43:20 +02:00
|
|
|
if (!xtrabackup_print_param)
|
|
|
|
usage();
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("mariabackup: Error: Please set parameter 'datadir'");
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Expand target-dir, incremental-basedir, etc. */
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
char cwd[FN_REFLEN];
|
2017-04-18 20:43:20 +02:00
|
|
|
my_getwd(cwd, sizeof(cwd), MYF(0));
|
|
|
|
|
|
|
|
my_load_path(xtrabackup_real_target_dir,
|
|
|
|
xtrabackup_target_dir, cwd);
|
|
|
|
unpack_dirname(xtrabackup_real_target_dir,
|
2017-04-18 21:05:57 +02:00
|
|
|
xtrabackup_real_target_dir);
|
2017-04-18 20:43:20 +02:00
|
|
|
xtrabackup_target_dir= xtrabackup_real_target_dir;
|
|
|
|
|
|
|
|
if (xtrabackup_incremental_basedir) {
|
|
|
|
my_load_path(xtrabackup_real_incremental_basedir,
|
|
|
|
xtrabackup_incremental_basedir, cwd);
|
|
|
|
unpack_dirname(xtrabackup_real_incremental_basedir,
|
2017-04-18 21:05:57 +02:00
|
|
|
xtrabackup_real_incremental_basedir);
|
2017-04-18 20:43:20 +02:00
|
|
|
xtrabackup_incremental_basedir =
|
|
|
|
xtrabackup_real_incremental_basedir;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_incremental_dir) {
|
|
|
|
my_load_path(xtrabackup_real_incremental_dir,
|
|
|
|
xtrabackup_incremental_dir, cwd);
|
|
|
|
unpack_dirname(xtrabackup_real_incremental_dir,
|
2017-04-18 21:05:57 +02:00
|
|
|
xtrabackup_real_incremental_dir);
|
2017-04-18 20:43:20 +02:00
|
|
|
xtrabackup_incremental_dir = xtrabackup_real_incremental_dir;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_extra_lsndir) {
|
|
|
|
my_load_path(xtrabackup_real_extra_lsndir,
|
|
|
|
xtrabackup_extra_lsndir, cwd);
|
|
|
|
unpack_dirname(xtrabackup_real_extra_lsndir,
|
2017-04-18 21:05:57 +02:00
|
|
|
xtrabackup_real_extra_lsndir);
|
2017-04-18 20:43:20 +02:00
|
|
|
xtrabackup_extra_lsndir = xtrabackup_real_extra_lsndir;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* get default temporary directory */
|
|
|
|
if (!opt_mysql_tmpdir || !opt_mysql_tmpdir[0]) {
|
|
|
|
opt_mysql_tmpdir = getenv("TMPDIR");
|
2021-06-06 13:21:03 +02:00
|
|
|
#if defined(_WIN32)
|
2017-04-18 20:43:20 +02:00
|
|
|
if (!opt_mysql_tmpdir) {
|
|
|
|
opt_mysql_tmpdir = getenv("TEMP");
|
|
|
|
}
|
|
|
|
if (!opt_mysql_tmpdir) {
|
|
|
|
opt_mysql_tmpdir = getenv("TMP");
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
if (!opt_mysql_tmpdir || !opt_mysql_tmpdir[0]) {
|
|
|
|
opt_mysql_tmpdir = const_cast<char*>(DEFAULT_TMPDIR);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* temporary setting of enough size */
|
|
|
|
srv_page_size_shift = UNIV_PAGE_SIZE_SHIFT_MAX;
|
|
|
|
srv_page_size = UNIV_PAGE_SIZE_MAX;
|
|
|
|
if (xtrabackup_backup && xtrabackup_incremental) {
|
|
|
|
/* direct specification is only for --backup */
|
|
|
|
/* and the lsn is prior to the other option */
|
|
|
|
|
|
|
|
char* endchar;
|
|
|
|
int error = 0;
|
|
|
|
incremental_lsn = strtoll(xtrabackup_incremental, &endchar, 10);
|
|
|
|
if (*endchar != '\0')
|
|
|
|
error = 1;
|
|
|
|
|
|
|
|
if (error) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: value '%s' may be wrong format for "
|
2019-01-14 22:28:23 +01:00
|
|
|
"incremental option.", xtrabackup_incremental);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
} else if (xtrabackup_backup && xtrabackup_incremental_basedir) {
|
|
|
|
char filename[FN_REFLEN];
|
|
|
|
|
|
|
|
sprintf(filename, "%s/%s", xtrabackup_incremental_basedir, XTRABACKUP_METADATA_FILENAME);
|
|
|
|
|
|
|
|
if (!xtrabackup_read_metadata(filename)) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: error: failed to read metadata from "
|
2019-01-14 22:28:23 +01:00
|
|
|
"%s", filename);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
incremental_lsn = metadata_to_lsn;
|
|
|
|
xtrabackup_incremental = xtrabackup_incremental_basedir; //dummy
|
|
|
|
} else if (xtrabackup_prepare && xtrabackup_incremental_dir) {
|
|
|
|
char filename[FN_REFLEN];
|
|
|
|
|
|
|
|
sprintf(filename, "%s/%s", xtrabackup_incremental_dir, XTRABACKUP_METADATA_FILENAME);
|
|
|
|
|
|
|
|
if (!xtrabackup_read_metadata(filename)) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: error: failed to read metadata from "
|
2019-01-14 22:28:23 +01:00
|
|
|
"%s", filename);
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
incremental_lsn = metadata_from_lsn;
|
|
|
|
incremental_to_lsn = metadata_to_lsn;
|
|
|
|
incremental_last_lsn = metadata_last_lsn;
|
|
|
|
xtrabackup_incremental = xtrabackup_incremental_dir; //dummy
|
|
|
|
|
|
|
|
} else if (opt_incremental_history_name) {
|
|
|
|
xtrabackup_incremental = opt_incremental_history_name;
|
|
|
|
} else if (opt_incremental_history_uuid) {
|
|
|
|
xtrabackup_incremental = opt_incremental_history_uuid;
|
|
|
|
} else {
|
|
|
|
xtrabackup_incremental = NULL;
|
|
|
|
}
|
|
|
|
|
2018-04-09 20:16:50 +02:00
|
|
|
if (xtrabackup_stream && !xtrabackup_backup) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("Warning: --stream parameter is ignored, it only works together with --backup.");
|
2018-04-09 20:16:50 +02:00
|
|
|
}
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
if (!xb_init()) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* --print-param */
|
|
|
|
if (xtrabackup_print_param) {
|
|
|
|
printf("%s", print_param_str.str().c_str());
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_SUCCESS);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
print_version();
|
|
|
|
if (xtrabackup_incremental) {
|
2019-01-14 22:28:23 +01:00
|
|
|
msg("incremental backup from " LSN_PF " is enabled.",
|
2017-04-18 20:43:20 +02:00
|
|
|
incremental_lsn);
|
|
|
|
}
|
|
|
|
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
if (xtrabackup_export && !srv_file_per_table) {
|
2017-10-23 12:37:28 +02:00
|
|
|
msg("mariabackup: auto-enabling --innodb-file-per-table due to "
|
2019-01-14 22:28:23 +01:00
|
|
|
"the --export option");
|
MDEV-21907: InnoDB: Enable -Wconversion on clang and GCC
The -Wconversion in GCC seems to be stricter than in clang.
GCC at least since version 4.4.7 issues truncation warnings for
assignments to bitfields, while clang 10 appears to only issue
warnings when the sizes in bytes rounded to the nearest integer
powers of 2 are different.
Before GCC 10.0.0, -Wconversion required more casts and would not
allow some operations, such as x<<=1 or x+=1 on a data type that
is narrower than int.
GCC 5 (but not GCC 4, GCC 6, or any later version) is complaining
about x|=y even when x and y are compatible types that are narrower
than int. Hence, we must rewrite some x|=y as
x=static_cast<byte>(x|y) or similar, or we must disable -Wconversion.
In GCC 6 and later, the warning for assigning wider to bitfields
that are narrower than 8, 16, or 32 bits can be suppressed by
applying a bitwise & with the exact bitmask of the bitfield.
For older GCC, we must disable -Wconversion for GCC 4 or 5 in such
cases.
The bitwise negation operator appears to promote short integers
to a wider type, and hence we must add explicit truncation casts
around them. Microsoft Visual C does not allow a static_cast to
truncate a constant, such as static_cast<byte>(1) truncating int.
Hence, we will use the constructor-style cast byte(~1) for such cases.
This has been tested at least with GCC 4.8.5, 5.4.0, 7.4.0, 9.2.1, 10.0.0,
clang 9.0.1, 10.0.0, and MSVC 14.22.27905 (Microsoft Visual Studio 2019)
on 64-bit and 32-bit targets (IA-32, AMD64, POWER 8, POWER 9, ARMv8).
2020-03-12 18:46:41 +01:00
|
|
|
srv_file_per_table = TRUE;
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/* cannot execute both for now */
|
|
|
|
{
|
|
|
|
int num = 0;
|
|
|
|
|
|
|
|
if (xtrabackup_backup) num++;
|
|
|
|
if (xtrabackup_prepare) num++;
|
|
|
|
if (xtrabackup_copy_back) num++;
|
|
|
|
if (xtrabackup_move_back) num++;
|
|
|
|
if (xtrabackup_decrypt_decompress) num++;
|
|
|
|
if (num != 1) { /* !XOR (for now) */
|
|
|
|
usage();
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-05-06 09:49:35 +02:00
|
|
|
ut_ad(!field_ref_zero);
|
|
|
|
if (auto b = aligned_malloc(UNIV_PAGE_SIZE_MAX, 4096)) {
|
|
|
|
field_ref_zero = static_cast<byte*>(
|
|
|
|
memset_aligned<4096>(b, 0, UNIV_PAGE_SIZE_MAX));
|
|
|
|
} else {
|
|
|
|
msg("Can't allocate memory for field_ref_zero");
|
|
|
|
return EXIT_FAILURE;
|
|
|
|
}
|
|
|
|
|
|
|
|
auto _ = make_scope_exit([]() {
|
|
|
|
aligned_free(const_cast<byte*>(field_ref_zero));
|
|
|
|
field_ref_zero = nullptr;
|
|
|
|
});
|
|
|
|
|
2017-04-18 20:43:20 +02:00
|
|
|
/* --backup */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (xtrabackup_backup && !xtrabackup_backup_func()) {
|
|
|
|
return(EXIT_FAILURE);
|
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
/* --prepare */
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
if (xtrabackup_prepare
|
|
|
|
&& !xtrabackup_prepare_func(argv)) {
|
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 21:05:57 +02:00
|
|
|
}
|
2017-04-18 20:43:20 +02:00
|
|
|
|
|
|
|
if (xtrabackup_copy_back || xtrabackup_move_back) {
|
|
|
|
if (!check_if_param_set("datadir")) {
|
2018-09-11 21:59:35 +02:00
|
|
|
mysql_data_home = get_default_datadir();
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
if (!copy_back())
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (xtrabackup_decrypt_decompress && !decrypt_decompress()) {
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_FAILURE);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
|
|
|
|
MDEV-12548 Initial implementation of Mariabackup for MariaDB 10.2
InnoDB I/O and buffer pool interfaces and the redo log format
have been changed between MariaDB 10.1 and 10.2, and the backup
code has to be adjusted accordingly.
The code has been simplified, and many memory leaks have been fixed.
Instead of the file name xtrabackup_logfile, the file name ib_logfile0
is being used for the copy of the redo log. Unnecessary InnoDB startup and
shutdown and some unnecessary threads have been removed.
Some help was provided by Vladislav Vaintroub.
Parameters have been cleaned up and aligned with those of MariaDB 10.2.
The --dbug option has been added, so that in debug builds,
--dbug=d,ib_log can be specified to enable diagnostic messages
for processing redo log entries.
By default, innodb_doublewrite=OFF, so that --prepare works faster.
If more crash-safety for --prepare is needed, double buffering
can be enabled.
The parameter innodb_log_checksums=OFF can be used to ignore redo log
checksums in --backup.
Some messages have been cleaned up.
Unless --export is specified, Mariabackup will not deal with undo log.
The InnoDB mini-transaction redo log is not only about user-level
transactions; it is actually about mini-transactions. To avoid confusion,
call it the redo log, not transaction log.
We disable any undo log processing in --prepare.
Because MariaDB 10.2 supports indexed virtual columns, the
undo log processing would need to be able to evaluate virtual column
expressions. To reduce the amount of code dependencies, we will not
process any undo log in prepare.
This means that the --export option must be disabled for now.
This also means that the following options are redundant
and have been removed:
xtrabackup --apply-log-only
innobackupex --redo-only
In addition to disabling any undo log processing, we will disable any
further changes to data pages during --prepare, including the change
buffer merge. This means that restoring incremental backups should
reliably work even when change buffering is being used on the server.
Because of this, preparing a backup will not generate any further
redo log, and the redo log file can be safely deleted. (If the
--export option is enabled in the future, it must generate redo log
when processing undo logs and buffered changes.)
In --prepare, we cannot easily know if a partial backup was used,
especially when restoring a series of incremental backups. So, we
simply warn about any missing files, and ignore the redo log for them.
FIXME: Enable the --export option.
FIXME: Improve the handling of the MLOG_INDEX_LOAD record, and write
a test that initiates a backup while an ALGORITHM=INPLACE operation
is creating indexes or rebuilding a table. An error should be detected
when preparing the backup.
FIXME: In --incremental --prepare, xtrabackup_apply_delta() should
ensure that if FSP_SIZE is modified, the file size will be adjusted
accordingly.
2017-06-30 09:49:37 +02:00
|
|
|
return(EXIT_SUCCESS);
|
2017-04-18 20:43:20 +02:00
|
|
|
}
|
2017-09-08 00:53:21 +02:00
|
|
|
|
|
|
|
|
|
|
|
static int get_exepath(char *buf, size_t size, const char *argv0)
|
|
|
|
{
|
|
|
|
#ifdef _WIN32
|
2018-02-06 13:55:58 +01:00
|
|
|
DWORD ret = GetModuleFileNameA(NULL, buf, (DWORD)size);
|
2017-09-08 00:53:21 +02:00
|
|
|
if (ret > 0)
|
|
|
|
return 0;
|
|
|
|
#elif defined(__linux__)
|
|
|
|
ssize_t ret = readlink("/proc/self/exe", buf, size-1);
|
|
|
|
if(ret > 0)
|
|
|
|
return 0;
|
2021-02-03 20:44:34 +01:00
|
|
|
#elif defined(__APPLE__)
|
|
|
|
size_t ret = proc_pidpath(getpid(), buf, static_cast<uint32_t>(size));
|
|
|
|
if (ret > 0) {
|
|
|
|
buf[ret] = 0;
|
|
|
|
return 0;
|
|
|
|
}
|
2021-03-19 10:31:27 +01:00
|
|
|
#elif defined(__FreeBSD__)
|
|
|
|
int mib[4] = {CTL_KERN, KERN_PROC, KERN_PROC_PATHNAME, -1};
|
|
|
|
if (sysctl(mib, 4, buf, &size, NULL, 0) == 0) {
|
|
|
|
return 0;
|
|
|
|
}
|
2017-09-08 00:53:21 +02:00
|
|
|
#endif
|
|
|
|
|
|
|
|
return my_realpath(buf, argv0, 0);
|
|
|
|
}
|
2018-12-29 02:06:19 +01:00
|
|
|
|
|
|
|
|
|
|
|
#if defined (__SANITIZE_ADDRESS__) && defined (__linux__)
|
|
|
|
/* Avoid LeakSanitizer's false positives. */
|
|
|
|
const char* __asan_default_options()
|
|
|
|
{
|
|
|
|
return "detect_leaks=0";
|
|
|
|
}
|
|
|
|
#endif
|