mariadb/storage/maria/ma_checkpoint.c

1188 lines
45 KiB
C
Raw Normal View History

WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* Copyright (C) 2006,2007 MySQL AB
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
GPL license update (same change as was done for all files in 5.1). storage/maria/Makefile.am: GPL license update storage/maria/ft_maria.c: GPL license update storage/maria/ha_maria.cc: GPL license update storage/maria/ha_maria.h: GPL license update storage/maria/lockman.c: GPL license update storage/maria/lockman.h: GPL license update storage/maria/ma_bitmap.c: GPL license update storage/maria/ma_blockrec.c: GPL license update storage/maria/ma_blockrec.h: GPL license update storage/maria/ma_cache.c: GPL license update storage/maria/ma_changed.c: GPL license update storage/maria/ma_check.c: GPL license update storage/maria/ma_checkpoint.c: GPL license update storage/maria/ma_checkpoint.h: GPL license update storage/maria/ma_checksum.c: GPL license update storage/maria/ma_close.c: GPL license update storage/maria/ma_control_file.c: GPL license update storage/maria/ma_control_file.h: GPL license update storage/maria/ma_create.c: GPL license update storage/maria/ma_dbug.c: GPL license update storage/maria/ma_delete.c: GPL license update storage/maria/ma_delete_all.c: GPL license update storage/maria/ma_delete_table.c: GPL license update storage/maria/ma_dynrec.c: GPL license update storage/maria/ma_extra.c: GPL license update storage/maria/ma_ft_boolean_search.c: GPL license update storage/maria/ma_ft_eval.c: GPL license update storage/maria/ma_ft_eval.h: GPL license update storage/maria/ma_ft_nlq_search.c: GPL license update storage/maria/ma_ft_parser.c: GPL license update storage/maria/ma_ft_stem.c: GPL license update storage/maria/ma_ft_test1.c: GPL license update storage/maria/ma_ft_test1.h: GPL license update storage/maria/ma_ft_update.c: GPL license update storage/maria/ma_ftdefs.h: GPL license update storage/maria/ma_fulltext.h: GPL license update storage/maria/ma_info.c: GPL license update storage/maria/ma_init.c: GPL license update storage/maria/ma_key.c: GPL license update storage/maria/ma_keycache.c: GPL license update storage/maria/ma_least_recently_dirtied.c: GPL license update storage/maria/ma_least_recently_dirtied.h: GPL license update storage/maria/ma_locking.c: GPL license update storage/maria/ma_open.c: GPL license update storage/maria/ma_packrec.c: GPL license update storage/maria/ma_page.c: GPL license update storage/maria/ma_panic.c: GPL license update storage/maria/ma_preload.c: GPL license update storage/maria/ma_range.c: GPL license update storage/maria/ma_recovery.c: GPL license update storage/maria/ma_recovery.h: GPL license update storage/maria/ma_rename.c: GPL license update storage/maria/ma_rfirst.c: GPL license update storage/maria/ma_rkey.c: GPL license update storage/maria/ma_rlast.c: GPL license update storage/maria/ma_rnext.c: GPL license update storage/maria/ma_rnext_same.c: GPL license update storage/maria/ma_rprev.c: GPL license update storage/maria/ma_rrnd.c: GPL license update storage/maria/ma_rsame.c: GPL license update storage/maria/ma_rsamepos.c: GPL license update storage/maria/ma_rt_index.c: GPL license update storage/maria/ma_rt_index.h: GPL license update storage/maria/ma_rt_key.c: GPL license update storage/maria/ma_rt_key.h: GPL license update storage/maria/ma_rt_mbr.c: GPL license update storage/maria/ma_rt_mbr.h: GPL license update storage/maria/ma_rt_split.c: GPL license update storage/maria/ma_rt_test.c: GPL license update storage/maria/ma_scan.c: GPL license update storage/maria/ma_search.c: GPL license update storage/maria/ma_sort.c: GPL license update storage/maria/ma_sp_defs.h: GPL license update storage/maria/ma_sp_key.c: GPL license update storage/maria/ma_sp_test.c: GPL license update storage/maria/ma_static.c: GPL license update storage/maria/ma_statrec.c: GPL license update storage/maria/ma_test1.c: GPL license update storage/maria/ma_test2.c: GPL license update storage/maria/ma_test3.c: GPL license update storage/maria/ma_unique.c: GPL license update storage/maria/ma_update.c: GPL license update storage/maria/ma_write.c: GPL license update storage/maria/maria_chk.c: GPL license update storage/maria/maria_def.h: GPL license update storage/maria/maria_ftdump.c: GPL license update storage/maria/maria_pack.c: GPL license update storage/maria/tablockman.c: GPL license update storage/maria/tablockman.h: GPL license update storage/maria/trnman.c: GPL license update storage/maria/trnman.h: GPL license update
2007-03-02 11:20:23 +01:00
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
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA */
/*
WL#3071 Maria checkpoint
First version written by Guilhem Bichot on 2006-04-27.
*/
/* Here is the implementation of this module */
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
/** @todo RECOVERY BUG this is unreviewed code */
/*
Summary:
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
checkpoints are done either by a background thread (checkpoint every Nth
second) or by a client.
In ha_maria, it's not made available to clients, and will soon be done by a
background thread (periodically taking checkpoints and flushing dirty
pages).
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
#include "maria_def.h"
#include "ma_pagecache.h"
#include "ma_blockrec.h"
#include "ma_checkpoint.h"
#include "ma_loghandler_lsn.h"
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/** @brief type of checkpoint currently running */
static CHECKPOINT_LEVEL checkpoint_in_progress= CHECKPOINT_NONE;
/** @brief protects checkpoint_in_progress */
static pthread_mutex_t LOCK_checkpoint;
/** @brief for killing the background checkpoint thread */
static pthread_cond_t COND_checkpoint;
/** @brief if checkpoint module was inited or not */
static my_bool checkpoint_inited= FALSE;
/** @brief 'kill' flag for the background checkpoint thread */
static int checkpoint_thread_die;
/* is ulong like pagecache->blocks_changed */
static ulong pages_to_flush_before_next_checkpoint;
static PAGECACHE_FILE *dfiles, /**< data files to flush in background */
*dfiles_end; /**< list of data files ends here */
static PAGECACHE_FILE *kfiles, /**< index files to flush in background */
*kfiles_end; /**< list of index files ends here */
/* those two statistics below could serve in SHOW GLOBAL STATUS */
static uint checkpoints_total= 0, /**< all checkpoint requests made */
checkpoints_ok_total= 0; /**< all checkpoints which succeeded */
struct st_filter_param
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
LSN up_to_lsn; /**< only pages with rec_lsn < this LSN */
uint max_pages; /**< stop after flushing this number pages */
}; /**< information to determine which dirty pages should be flushed */
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
static enum pagecache_flush_filter_result
filter_flush_file_medium(enum pagecache_page_type type,
pgcache_page_no_t page,
LSN rec_lsn, void *arg);
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
static enum pagecache_flush_filter_result
filter_flush_file_full(enum pagecache_page_type type,
pgcache_page_no_t page,
LSN rec_lsn, void *arg);
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
static enum pagecache_flush_filter_result
filter_flush_file_evenly(enum pagecache_page_type type,
pgcache_page_no_t pageno,
LSN rec_lsn, void *arg);
static int really_execute_checkpoint(void);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pthread_handler_t ma_checkpoint_background(void *arg);
static int collect_tables(LEX_STRING *str, LSN checkpoint_start_log_horizon);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/**
@brief Does a checkpoint
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
@param level what level of checkpoint to do
@param no_wait if another checkpoint of same or stronger level
is already running, consider our job done
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
@note In ha_maria, there can never be two threads trying a checkpoint at
the same time.
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
@return Operation status
@retval 0 ok
@retval !=0 error
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
int ma_checkpoint_execute(CHECKPOINT_LEVEL level, my_bool no_wait)
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
int result= 0;
DBUG_ENTER("ma_checkpoint_execute");
WL#3072 Maria Recovery * recovery from ha_maria now skips replaying DDLs (too dangerous) * maria_read_log still replays DDLs, print warning about issues * fixes to replaying of REDO_RENAME * don't replay DDLs on corrupted tables (safer) * print a one-line message when really doing a recovery (applies to ha_maria, not maria_read_log) i.e. some REDOs or UNDOs are read. storage/maria/ma_checkpoint.c: fix for assertion failure storage/maria/ma_recovery.c: * Recovery from ha_maria now skips replaying DDLs (as the initial plan said) as this is unsafe in case of crashes during the DDL; applying the records may do harm (destroy important files) so we prefer to leave the "mess" of files untouched. A proper recovery of DDLs requires very careful thinking, probably testing separately the existence of the data and index file instead of using maria_open() which tests the existence of both, and maybe storing create_rename_lsn in the data file too. * maria_read_log still replays DDLs, we print a warning about dangers (due to ALTER TABLE not logging insertions into the tmp table; we will maybe need an option to have logging of those insertions). * fixes to replaying of REDO_RENAME (test create_rename_lsn of 'new_name' table if it exists; if that table exists and is more recent than the record, remove the 'old_name' table). * don't replay DDLs on corrupted tables (play safe) * fail also in non-debug builds if table is open when it should not be (when creating it for example, it should not be already open). * when the trace file is not stdout (i.e. when this is ha_maria), if really doing a recovery (reading REDOs or UNDOs), print a one-line message to stderr to inform about start and end of recovery (useful to know what mysqld is doing, especially if it takes long or crashes). storage/maria/ma_recovery.h: parameter to replay DDLs or not storage/maria/maria_read_log.c: replay DDLs in maria_read_log, to be able to recreate tables from scratch.
2007-09-15 14:45:26 +02:00
if (!checkpoint_inited)
{
/*
If ha_maria failed to start, maria_panic_hton is called, we come here.
*/
DBUG_RETURN(0);
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
DBUG_ASSERT(level > CHECKPOINT_NONE);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* look for already running checkpoints */
pthread_mutex_lock(&LOCK_checkpoint);
while (checkpoint_in_progress != CHECKPOINT_NONE)
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
if (no_wait && (checkpoint_in_progress >= level))
{
/*
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
If we are the checkpoint background thread, we don't wait (it's
smarter to flush pages instead of waiting here while the other thread
finishes its checkpoint).
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pthread_mutex_unlock(&LOCK_checkpoint);
goto end;
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pthread_cond_wait(&COND_checkpoint, &LOCK_checkpoint);
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
checkpoint_in_progress= level;
pthread_mutex_unlock(&LOCK_checkpoint);
/* from then on, we are sure to be and stay the only checkpointer */
result= really_execute_checkpoint();
pthread_cond_broadcast(&COND_checkpoint);
end:
2006-12-18 17:24:02 +01:00
DBUG_RETURN(result);
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/**
@brief Does a checkpoint, really; expects no other checkpoints
running.
Checkpoint level requested is read from checkpoint_in_progress.
@return Operation status
@retval 0 ok
@retval !=0 error
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
static int really_execute_checkpoint(void)
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
uint i, error= 0;
/** @brief checkpoint_start_log_horizon will be stored there */
char *ptr;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
LEX_STRING record_pieces[4]; /**< only malloc-ed pieces */
LSN min_page_rec_lsn, min_trn_rec_lsn, min_first_undo_lsn;
TRANSLOG_ADDRESS checkpoint_start_log_horizon;
char checkpoint_start_log_horizon_char[LSN_STORE_SIZE];
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
DBUG_ENTER("really_execute_checkpoint");
WL#3072 - Maria Recovery * to honour WAL we now force the whole log when flushing a bitmap page. * ability to intentionally crash in various places for recovery testing * bugfix (dirty pages list found in checkpoint record was ignored) * smaller checkpoint record * misc small cleanups and comments mysql-test/include/maria_empty_logs.inc: maria-purge.test creates ~11 logs, remove them all mysql-test/r/maria-recovery-bitmap.result: result is good; without the _ma_bitmap_get_log_address() call, we got check error Bitmap at 0 has pages reserved outside of data file length mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery-bitmap.test: enable test of "bitmap-flush should flush whole log otherwise corrupted data file (bitmap ahead of data pages)". mysql-test/t/maria-recovery.test: test of checkpoint sql/sql_table.cc: comment storage/maria/ha_maria.cc: _ma_reenable_logging_for_table() now includes file->trn=0. At the end of repair() we don't need to re-enable logging, it is done already by caller (like copy_data_between_tables()); it sounds strange that this function could decide to re-enable, it should be up to caller who knows what other operations it plans. Removing this line led to assertion failure in maria_lock_database(F_UNLCK), fixed by removing the assertion: maria_lock_database() is here called in a context where F_UNLCK does not make the table visible to others so assertion is excessive, and external_lock() is already designed to honour the asserted condition. Ability to crash at the end of bulk insert when indices have been enabled. storage/maria/ma_bitmap.c: Better use pagecache_file_init() than set pagecache callbacks directly; and a new function to set those callbacks for bitmap so that we can reuse it. _ma_bitmap_get_log_address() is a pagecache get_log_address callback which causes the whole log to be flushed when a bitmap page is flushed by the page cache. This was required by WAL. storage/maria/ma_blockrec.c: get_log_address pagecache callback for data (non bitmap) pages: just reads the LSN from the page's content, like was hard-coded before in ma_pagecache.c. storage/maria/ma_blockrec.h: functions which need to be exported storage/maria/ma_check.c: create_new_data_handle() can be static. Ability to crash after rebuilding the index in OPTIMIZE, in REPAIR. my_lock() implemented already. storage/maria/ma_checkpoint.c: As MARIA_SHARE* is now accessible to pagecache_collect_changed_blocks_LSN(), we don't need to store kfile/dfile descriptors in checkpoint record, 2-byte-id of the table plus one byte to say if this is data or index file is enough. So we go from 4+4 bytes per table down to 2+1. storage/maria/ma_commit.c: removing duplicate functions (see _ma_tmp_disable_logging_for_table()) storage/maria/ma_extra.c: Monty fixed storage/maria/ma_key_recover.c: comment storage/maria/ma_locking.c: Sometimes other code does funny things with maria_lock_database(), like ha_maria::repair() calling it at start and end without going through ha_maria::external_lock(). So it happens that maria_lock_database() is called with now_transactional!=born_transactional. storage/maria/ma_loghandler.c: update to new prototype storage/maria/ma_open.c: set_data|index_pagecache_callbacks() need to be exported as they are now called when disabling/enabling transactionality. storage/maria/ma_pagecache.c: Removing PAGE_LSN_OFFSET, as much of the code relies on it being 0 anyway (let's not give impression we can just change this constant). When flushing a page to disk, call the get_log_address callback to know up to which LSN the log should be flushed. As we now can access MARIA_SHARE* we can know share->id and store it into the checkpoint record; we thus go from 4 bytes per dirty page to 2+1. storage/maria/ma_pagecache.h: get_log_address callback storage/maria/ma_panic.c: No reason to reset pagecache callbacks in HA_PANIC_READ: all we do is reopen files if they were closed; callbacks should be in place already as 'info' exists; we just want to modify the file descriptors, not the full PAGECACHE_FILE structure. If we open data file and it was closed, share->bitmap.file needs to be set. Note that the modified code is disabled anyway. storage/maria/ma_recovery.c: Checkpoint record does not contain kfile/dfile descriptors anymore so code can be simplified. Hash key in all_dirty_pages is not made from file_descriptor & pageno anymore, but index_or_data & table-short-id & pageno. If a table's create_rename_lsn is higher than record's LSN, we skip the table and don't fail if it's corrupted (because the LSNs say that we don't have to look at this table). If a table is skipped (for example due to create_rename_lsn), its UNDOs still cause undo_lsn to advance; this is so that if later we notice the transaction has to rollback we fail (as table should not be skipped in this case). Fixing a bug: the dirty_pages list was never used, because the LSN below which it was used was the minimum rec_lsn of dirty pages! It is now the min(checkpoint_start_log_horizon, min(trn's rec_lsn)). When we disable/reenable transactionality, we modify pagecache callbacks (needed for example for get_log_address: changing share->page_type is not enough anymore). storage/maria/ma_write.c: 'records' and 'checksum' are protected: they are updated under log's mutex in write-hooks when UNDO is written. storage/maria/maria_chk.c: remove use of duplicate functions. storage/maria/maria_def.h: set_data|index_pagecache_callbacks() need to be exported; _ma_reenable_logging_for_table() changes to a real function. storage/maria/unittest/ma_pagecache_consist.c: new prototype storage/maria/unittest/ma_pagecache_single.c: new prototype storage/maria/unittest/ma_test_loghandler_pagecache-t.c: new prototype
2007-12-30 21:32:07 +01:00
DBUG_PRINT("enter", ("level: %d", checkpoint_in_progress));
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
bzero(&record_pieces, sizeof(record_pieces));
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/*
STEP 1: record current end-of-log position using log's lock. It is
critical for the correctness of Checkpoint (related to memory visibility
rules, the log's lock is a mutex).
"Horizon" is a lower bound of the LSN of the next log record.
*/
checkpoint_start_log_horizon= translog_get_horizon();
DBUG_PRINT("info",("checkpoint_start_log_horizon (%lu,0x%lx)",
LSN_IN_PARTS(checkpoint_start_log_horizon)));
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
lsn_store(checkpoint_start_log_horizon_char, checkpoint_start_log_horizon);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/*
STEP 2: fetch information about transactions.
We must fetch transactions before dirty pages. Indeed, a transaction
first sets its rec_lsn then sets the page's rec_lsn then sets its rec_lsn
to 0. If we fetched pages first, we may see no dirty page yet, then we
fetch transactions but the transaction has already reset its rec_lsn to 0
so we miss rec_lsn again.
For a similar reason (over-allocated bitmap pages) we have to fetch
transactions before flushing bitmap pages.
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
min_trn_rec_lsn will serve to lower the starting point of the REDO phase
(down from checkpoint_start_log_horizon).
*/
if (unlikely(trnman_collect_transactions(&record_pieces[0],
&record_pieces[1],
&min_trn_rec_lsn,
&min_first_undo_lsn)))
goto err;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* STEP 3: fetch information about table files */
if (unlikely(collect_tables(&record_pieces[2],
checkpoint_start_log_horizon)))
2006-12-18 17:24:02 +01:00
goto err;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* STEP 4: fetch information about dirty pages */
/*
It's better to do it _after_ having flushed some data pages (which
collect_tables() may have done), because those are now non-dirty and so we
have a more up-to-date dirty pages list to put into the checkpoint record,
and thus we will have less work at Recovery.
*/
/* Using default pagecache for now */
if (unlikely(pagecache_collect_changed_blocks_with_lsn(maria_pagecache,
&record_pieces[3],
&min_page_rec_lsn)))
goto err;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* LAST STEP: now write the checkpoint log record */
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
LSN lsn;
Added --loose-skip-maria to MYSQLD_BOOTSTRAP_CMD to get bootstrap.test to work Allow one to run bootstrap even if --skip-maria is used (needed for bootstrap.test) Fixed lots of compiler warnings NOTE: maria-big and maria-recover tests failes becasue of bugs in transaction log handling. Sanja knows about this and is working on it! mysql-test/mysql-test-run.pl: Added --loose-skip-maria to MYSQLD_BOOTSTRAP_CMD to get bootstrap.test to work mysql-test/r/maria-recovery.result: Updated results mysql-test/t/bootstrap.test: Removed not needed empty line mysql-test/t/change_user.test: Fixed results for 32 bit systems mysql-test/t/maria-big.test: Only run this when you use --big mysql-test/t/maria-recovery.test: Added test case for recovery with big blobs mysys/my_uuid.c: Fixed compiler warning sql/mysqld.cc: Allow one to run bootstrap even if --skip-maria is used (needed for bootstrap.test) sql/set_var.cc: Compare max_join_size with ULONG_MAX instead of HA_POS_ERROR as we set max_join_size to ULONG_MAX by default storage/maria/ma_bitmap.c: Added __attribute((unused)) to fix compiler warning storage/maria/ma_blockrec.c: Added casts to remove compiler warnings Change variable types to avoid compiler warnings storage/maria/ma_check.c: Added casts to remove compiler warnings storage/maria/ma_checkpoint.c: Change variable types to avoid compiler warnings storage/maria/ma_create.c: Change variable types to avoid compiler warnings storage/maria/ma_delete.c: Added casts to remove compiler warnings storage/maria/ma_key_recover.c: Added casts to remove compiler warnings storage/maria/ma_loghandler.c: Moved initiazation of prev_buffer first as this could otherwise not be set in case of errors storage/maria/ma_page.c: Added casts to remove compiler warnings storage/maria/ma_pagecache.c: Added __attribute((unused)) to fix compiler warning storage/maria/ma_pagecrc.c: Added #ifndef DBUG_OFF to remove compiler warning storage/maria/ma_recovery.c: Added casts to remove compiler warnings storage/maria/ma_write.c: Added casts to remove compiler warnings storage/maria/maria_chk.c: Split long string into two to avoid compiler warnings storage/myisam/ft_boolean_search.c: Added LINT_INIT() to remove compiler warning support-files/compiler_warnings.supp: Suppress wrong compiler warning unittest/mytap/tap.c: Fixed declaration to match prototypes to remove compiler warnings
2008-01-11 00:47:52 +01:00
translog_size_t total_rec_length;
/*
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
the log handler is allowed to modify "str" and "length" (but not "*str")
of its argument, so we must not pass it record_pieces directly,
otherwise we would later not know what memory pieces to my_free().
*/
Injecting more "const" declarations into code which does not change pointed data. I ran gcc -Wcast-qual on storage/maria, this identified un-needed casts, a couple of functions which said they had a const parameter though they changed the pointed content! This is fixed here. Some suspicious places receive a comment. The original intention of running -Wcast-qual was to find what code changes R-tree keys: I added const words, but hidden casts like those of int2store (casts target to (uint16*)) removed const checking; -Wcast-qual helped find those hidden casts. Log handler does not change the content pointed by LEX_STRING::str it receives, so we now use a struct which has a const inside, to emphasize this and be able to pass "const uchar*" buffers to log handler without fear of their content being changed by it. One-line fix for a merge glitch (when merging from MyISAM). include/m_string.h: As Maria's log handler uses LEX_STRING but never changes the content pointed by LEX_STRING::str, and assigns uchar* into this member most of the time, we introduce a new struct LEX_CUSTRING (C const U unsigned) for the log handler. include/my_global.h: In macros which read pointed content: use const pointers so that gcc -Wcast-qual does not warn about casting a const pointer to non-const. include/my_handler.h: In macros which read pointed content: use const pointers so that gcc -Wcast-qual does not warn about casting a const pointer to non-const. ha_find_null() does not change *a. include/my_sys.h: insert_dynamic() does not change *element. include/myisampack.h: In macros which read pointed content: use const pointers so that gcc -Wcast-qual does not warn about casting a const pointer to non-const. mysys/array.c: insert_dynamic() does not change *element mysys/my_handler.c: ha_find_null() does not change *a storage/maria/ma_bitmap.c: Log handler receives const strings now storage/maria/ma_blockrec.c: Log handler receives const strings now. _ma_apply_undo_row_delete/update() do change *header. storage/maria/ma_blockrec.h: correct prototype storage/maria/ma_check.c: Log handler receives const strings now. Un-needed casts storage/maria/ma_checkpoint.c: Log handler receives const strings now storage/maria/ma_checksum.c: unneeded cast storage/maria/ma_commit.c: Log handler receives const strings now storage/maria/ma_create.c: Log handler receives const strings now storage/maria/ma_dbug.c: fixing warning of gcc -Wcast-qual storage/maria/ma_delete.c: Log handler receives const strings now storage/maria/ma_delete_all.c: Log handler receives const strings now storage/maria/ma_delete_table.c: Log handler receives const strings now storage/maria/ma_dynrec.c: fixing some warnings of gcc -Wcast-qual. Unneeded casts removed. Comment about function which lies. storage/maria/ma_ft_parser.c: fix for warnings of gcc -Wcast-qual, removing unneeded casts storage/maria/ma_ft_update.c: less casts, comment storage/maria/ma_key.c: less casts, stay const (warnings of gcc -Wcast-qual) storage/maria/ma_key_recover.c: Log handler receives const strings now storage/maria/ma_loghandler.c: Log handler receives const strings now storage/maria/ma_loghandler.h: Log handler receives const strings now storage/maria/ma_loghandler_lsn.h: In macros which read pointed content: use const pointers so that gcc -Wcast-qual does not warn about casting a const pointer to non-const. storage/maria/ma_page.c: Log handler receives const strings now; more const storage/maria/ma_recovery.c: Log handler receives const strings now storage/maria/ma_rename.c: Log handler receives const strings now storage/maria/ma_rt_index.c: more const, to emphasize that functions don't change pointed content. best_key= NULL was forgotten during merge from MyISAM a few days ago, was causing a Valgrind warning storage/maria/ma_rt_index.h: new proto storage/maria/ma_rt_key.c: more const storage/maria/ma_rt_key.h: new proto storage/maria/ma_rt_mbr.c: more const for functions which deserve it storage/maria/ma_rt_mbr.h: new prototype storage/maria/ma_rt_split.c: make const what is not changed. storage/maria/ma_search.c: un-needed casts, more const storage/maria/ma_sp_key.c: more const storage/maria/ma_unique.c: un-needed casts. storage/maria/ma_write.c: Log handler receives const strings now storage/maria/maria_def.h: some more const storage/maria/unittest/ma_test_loghandler-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_first_lsn-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_max_lsn-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_multigroup-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_multithread-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_noflush-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_nologs-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_pagecache-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_purge-t.c: Log handler receives const strings now
2008-04-03 15:40:25 +02:00
LEX_CUSTRING log_array[TRANSLOG_INTERNAL_PARTS + 5];
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
log_array[TRANSLOG_INTERNAL_PARTS + 0].str=
checkpoint_start_log_horizon_char;
log_array[TRANSLOG_INTERNAL_PARTS + 0].length= total_rec_length=
sizeof(checkpoint_start_log_horizon_char);
for (i= 0; i < (sizeof(record_pieces)/sizeof(record_pieces[0])); i++)
{
Injecting more "const" declarations into code which does not change pointed data. I ran gcc -Wcast-qual on storage/maria, this identified un-needed casts, a couple of functions which said they had a const parameter though they changed the pointed content! This is fixed here. Some suspicious places receive a comment. The original intention of running -Wcast-qual was to find what code changes R-tree keys: I added const words, but hidden casts like those of int2store (casts target to (uint16*)) removed const checking; -Wcast-qual helped find those hidden casts. Log handler does not change the content pointed by LEX_STRING::str it receives, so we now use a struct which has a const inside, to emphasize this and be able to pass "const uchar*" buffers to log handler without fear of their content being changed by it. One-line fix for a merge glitch (when merging from MyISAM). include/m_string.h: As Maria's log handler uses LEX_STRING but never changes the content pointed by LEX_STRING::str, and assigns uchar* into this member most of the time, we introduce a new struct LEX_CUSTRING (C const U unsigned) for the log handler. include/my_global.h: In macros which read pointed content: use const pointers so that gcc -Wcast-qual does not warn about casting a const pointer to non-const. include/my_handler.h: In macros which read pointed content: use const pointers so that gcc -Wcast-qual does not warn about casting a const pointer to non-const. ha_find_null() does not change *a. include/my_sys.h: insert_dynamic() does not change *element. include/myisampack.h: In macros which read pointed content: use const pointers so that gcc -Wcast-qual does not warn about casting a const pointer to non-const. mysys/array.c: insert_dynamic() does not change *element mysys/my_handler.c: ha_find_null() does not change *a storage/maria/ma_bitmap.c: Log handler receives const strings now storage/maria/ma_blockrec.c: Log handler receives const strings now. _ma_apply_undo_row_delete/update() do change *header. storage/maria/ma_blockrec.h: correct prototype storage/maria/ma_check.c: Log handler receives const strings now. Un-needed casts storage/maria/ma_checkpoint.c: Log handler receives const strings now storage/maria/ma_checksum.c: unneeded cast storage/maria/ma_commit.c: Log handler receives const strings now storage/maria/ma_create.c: Log handler receives const strings now storage/maria/ma_dbug.c: fixing warning of gcc -Wcast-qual storage/maria/ma_delete.c: Log handler receives const strings now storage/maria/ma_delete_all.c: Log handler receives const strings now storage/maria/ma_delete_table.c: Log handler receives const strings now storage/maria/ma_dynrec.c: fixing some warnings of gcc -Wcast-qual. Unneeded casts removed. Comment about function which lies. storage/maria/ma_ft_parser.c: fix for warnings of gcc -Wcast-qual, removing unneeded casts storage/maria/ma_ft_update.c: less casts, comment storage/maria/ma_key.c: less casts, stay const (warnings of gcc -Wcast-qual) storage/maria/ma_key_recover.c: Log handler receives const strings now storage/maria/ma_loghandler.c: Log handler receives const strings now storage/maria/ma_loghandler.h: Log handler receives const strings now storage/maria/ma_loghandler_lsn.h: In macros which read pointed content: use const pointers so that gcc -Wcast-qual does not warn about casting a const pointer to non-const. storage/maria/ma_page.c: Log handler receives const strings now; more const storage/maria/ma_recovery.c: Log handler receives const strings now storage/maria/ma_rename.c: Log handler receives const strings now storage/maria/ma_rt_index.c: more const, to emphasize that functions don't change pointed content. best_key= NULL was forgotten during merge from MyISAM a few days ago, was causing a Valgrind warning storage/maria/ma_rt_index.h: new proto storage/maria/ma_rt_key.c: more const storage/maria/ma_rt_key.h: new proto storage/maria/ma_rt_mbr.c: more const for functions which deserve it storage/maria/ma_rt_mbr.h: new prototype storage/maria/ma_rt_split.c: make const what is not changed. storage/maria/ma_search.c: un-needed casts, more const storage/maria/ma_sp_key.c: more const storage/maria/ma_unique.c: un-needed casts. storage/maria/ma_write.c: Log handler receives const strings now storage/maria/maria_def.h: some more const storage/maria/unittest/ma_test_loghandler-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_first_lsn-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_max_lsn-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_multigroup-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_multithread-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_noflush-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_nologs-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_pagecache-t.c: Log handler receives const strings now storage/maria/unittest/ma_test_loghandler_purge-t.c: Log handler receives const strings now
2008-04-03 15:40:25 +02:00
log_array[TRANSLOG_INTERNAL_PARTS + 1 + i]=
*(LEX_CUSTRING *)&record_pieces[i];
Fixed compiler warnings by adding casts and changing variable types Fixed bug that caused change_user.test to fail Fixed bug that caused mysql_client_test to fail include/myisam.h: Fixed prototypes mysql-test/r/create.result: Fix that test works even if Maria is not used for temporary tables mysql-test/t/create.test: Fix that test works even if Maria is not used for temporary tables sql/mysqld.cc: Fixed that default value of max_join_size is set correctly; It needs to match usage in set_var.cc sql/set_var.cc: Fixed test, now when max_join_size is initialized correctly sql/sql_select.cc: Fixed that one can compile without -DUSE_MARIA_FOR_TMP_TABLES storage/maria/ma_blockrec.c: Fixed compiler warnings by adding casts storage/maria/ma_checkpoint.c: Fixed compiler warnings by adding casts storage/maria/ma_create.c: Fixed compiler warnings by adding casts storage/maria/ma_delete_table.c: Fixed compiler warnings by adding casts storage/maria/ma_loghandler.c: Fixed compiler warnings by adding casts and changing types for variables Changed translog_new_page_header to use changing integer instead of calling time() as time() is a slow call and will give same results when calling many times withing one second storage/maria/ma_pagecrc.c: Fixed compiler warnings by adding casts storage/maria/ma_recovery.c: Fixed indentation storage/myisam/ha_myisam.cc: Fixed wrong types for variables Changed chk_data_link() and repair*() functions to take my_bool as argument storage/myisam/mi_check.c: Fixes to handle that param.test_flag is now ulonglong storage/myisam/myisamchk.c: Fixes to handle that param.test_flag is now ulonglong support-files/compiler_warnings.supp: Fixed line numbers
2008-01-11 18:39:43 +01:00
total_rec_length+= (translog_size_t) record_pieces[i].length;
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
if (unlikely(translog_write_record(&lsn, LOGREC_CHECKPOINT,
&dummy_transaction_object, NULL,
total_rec_length,
sizeof(log_array)/sizeof(log_array[0]),
WL#3072 - Maria recovery. * Recovery of the table's live checksum (CREATE TABLE ... CHECKSUM=1) is achieved in this patch. The table's live checksum (info->s->state.state.checksum) is updated in inwrite_rec_hook's under the log mutex when writing UNDO_ROW_INSERT|UPDATE|DELETE and REDO_DELETE_ALL. The checksum variation caused by the operation is stored in these UNDOs, so that the REDO phase, when it sees such UNDOs, can update the live checksum if it is older (state.is_of_lsn is lower) than the record. It is also used, as a nice add-on with no cost, to do less row checksum computation during the UNDO phase (as we have it in the record already). Doing this work, it became pressing to move in-write hooks (write_hook_for_redo() et al) to ma_blockrec.c. The 'parts' argument of inwrite_rec_hook is unpredictable (it comes mangled at this stage, for example by LSN compression) so it is replaced by a 'void* hook_arg', which is used to pass down information, currently only to write_hook_for_clr_end() (previous undo_lsn and type of undone record). * If from ha_maria, we print to stderr how many seconds (with one fractional digit) the REDO phase took, same for UNDO phase and for final table close. Just to give an indication for debugging and maybe also for Support. storage/maria/ha_maria.cc: question for Monty storage/maria/ma_blockrec.c: * log in-write hooks (write_hook_for_redo() etc) move from ma_loghandler.c to here; this is natural: the hooks are coupled to their callers (functions in ma_blockrec.c). * translog_write_record() now has a new argument "hook_arg"; using it to pass down to write_hook_for_clr_end() the transaction's previous_undo_lsn and the type of the being undone record, and also to pass down to all UNDOs the live checksum variation caused by the operation. * If table has live checksum, store in UNDO_ROW_INSERT|UPDATE|DELETE and in CLR_END the checksum variation ("delta") caused by the operation. For example if a DELETE caused the table's live checksum to change from 123 to 456, we store in the UNDO_ROW_DELETE, in 4 bytes, the value 333 (456-123). * Instead of hard-coded "1" as length of the place where we store the undone record's type in CLR_END, use a symbol CLR_TYPE_STORE_SIZE; use macros clr_type_store and clr_type_korr. * write_block_record() has a new parameter 'old_record_checksum' which is the pre-computed checksum of old_record; that value is used to update the table's live checksum when writing UNDO_ROW_UPDATE|CLR_END. * In allocate_write_block_record(), if we are executing UNDO_ROW_DELETE the row's checksum is already computed. * _ma_update_block_record2() now expect the new row's checksum into cur_row.checksum (was already true) and the old row's checksum into new_row.checksum (that's new). Its two callers, maria_update() and _ma_apply_undo_row_update(), honour this. * When executing an UNDO_ROW_INSERT|UPDATE|DELETE in UNDO phase, pick up the checksum delta from the log record. It is then used to update the table's live checksum when writing CLR_END, and saves us a computation of record. storage/maria/ma_blockrec.h: in-write hooks move from ma_loghandler.c storage/maria/ma_check.c: more straightforward size of buffer storage/maria/ma_checkpoint.c: <= is enough storage/maria/ma_commit.c: new prototype of translog_write_record() storage/maria/ma_create.c: new prototype of translog_write_record() storage/maria/ma_delete.c: The row's checksum must be computed before calling(*delete_record)(), not after, because it must be known inside _ma_delete_block_record() (to update the table's live checksum when writing UNDO_ROW_DELETE). If deleting from a transactional table, live checksum was already updated when writing UNDO_ROW_DELETE. storage/maria/ma_delete_all.c: @todo is now done (in ma_loghandler.c) storage/maria/ma_delete_table.c: new prototype of translog_write_record() storage/maria/ma_loghandler.c: * in-write hooks move to ma_blockrec.c. * translog_write_record() gets a new argument 'hook_arg' which is passed down to pre|inwrite_rec_hook. It is more useful that 'parts' for those hooks, because when those hooks are called, 'parts' has possibly been mangled (like with LSN compression) and is so unpredictable. * fix for compiler warning (unused buffer_start when compiling without debug support) * Because checksum delta is stored into UNDO_ROW_INSERT|UPDATE|DELETE and CLR_END, but only if the table has live checksum, these records are not PSEUDOFIXEDLENGTH anymore, they are now VARIABLE_LENGTH (their length is X if no live checksum and X+4 otherwise). * add an inwrite_rec_hook for UNDO_ROW_UPDATE, which updates the table's live checksum. Update it also in hooks of UNDO_ROW_INSERT| DELETE and REDO_DELETE_ALL and CLR_END. * Bugfix: when reading a record in translog_read_record(), it happened that "length" became negative, because the function assumed that the record extended beyond the page's end, whereas it may be shorter. storage/maria/ma_loghandler.h: * Instead of hard-coded "1" and "4", use symbols and macros to store/retrieve the type of record which the CLR_END corresponds to, and the checksum variation caused by the operation which logs the record * translog_write_record() gets a new argument 'hook_arg' which is passed down to pre|inwrite_rec_hook. It is more useful that 'parts' for those hooks, because when those hooks are called, 'parts' has possibly been mangled (like with LSN compression) and is so unpredictable. storage/maria/ma_open.c: fix for "empty body in if() statement" (when compiling without safemutex) storage/maria/ma_pagecache.c: <= is enough storage/maria/ma_recovery.c: * print the time that each recovery phase (REDO/UNDO/flush) took; this is enabled only when recovering from ha_maria. Is it printed n seconds with a fractional part of one digit (like 123.4 seconds). * In the REDO phase, update the table's live checksum by using the checksum delta stored in UNDO_ROW_INSERT|DELETE|UPDATE and CLR_END. Update it too when seeing REDO_DELETE_ALL. * In the UNDO phase, when executing UNDO_ROW_INSERT, if the table does not have live checksum then reading the record's header (as done by the master loop of run_undo_phase()) is enough; otherwise we do a translog_read_record() to have the checksum delta ready for _ma_apply_undo_row_insert(). * When at the end of the REDO phase we notice that there is an unfinished group of REDOs, don't assert in debug binaries, as I verified that it can happen in real life (with kill -9) * removing ' in #error as it confuses gcc3 storage/maria/ma_rename.c: new prototype of translog_write_record() storage/maria/ma_test_recovery.expected: Change in output of ma_test_recovery: now all live checksums of original tables equal those of tables recreated by the REDO phase and those of tables fixed by the UNDO phase. I.e. recovery of the live checksum looks like working (which was after all the only goal of this changeset). I checked by hand that it's not just all live checksums which are now 0 and that's why they match. They are the old values like 3757530372. maria.test has hard-coded checksum values in its result file so checks this too. storage/maria/ma_update.c: * It's useless to put up HA_STATE_CHANGED in 'key_changed', as we put up HA_STATE_CHANGED in info->update anyway. * We need to compute the old and new rows' checksum before calling (*update_record)(), as checksum delta must be known when logging UNDO_ROW_UPDATE which is done by _ma_update_block_record(). Note that some functions change the 'newrec' record (at least _ma_check_unique() does) so we cannot move the checksum computation too early in the function. storage/maria/ma_write.c: If inserting into a transactional table, live's checksum was already updated when writing UNDO_ROW_INSERT. The multiplication is a trick to save an if(). storage/maria/unittest/ma_test_loghandler-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_first_lsn-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_max_lsn-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_multigroup-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_multithread-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_noflush-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_pagecache-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_purge-t.c: new prototype of translog_write_record() storage/myisam/sort.c: fix for compiler warnings in pushbuild (write_merge_key* functions didn't have their declaration match MARIA_HA::write_key).
2007-10-02 18:02:09 +02:00
log_array, NULL, NULL) ||
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
translog_flush(lsn)))
goto err;
translog_lock();
/*
This cannot be done as a inwrite_rec_hook of LOGREC_CHECKPOINT, because
such hook would be called before translog_flush (and we must be sure
that log was flushed before we write to the control file).
*/
Store maximum transaction id into control file at clean shutdown. This can serve to maria_chk to check that trids found in rows and keys are not too big. Also used by Recovery when logs are lost. Options --require-control-file, --datadir, --log-dir (yes, the dashes are inconsistent but I imitated mysqld --datadir and --maria-log-dir) for maria_chk. Lock control file _before_ reading its content. storage/maria/ha_maria.cc: new prototype storage/maria/ma_check.c: A function to find the max trid in the system (consults transaction manager and control file), to check tables. storage/maria/ma_checkpoint.c: new prototype storage/maria/ma_control_file.c: Store max trid into control file, in a backward-compatible way (can still read old control files). Parameter to ma_control_file_open(), to not create the log if it's missing (maria_chk needs that). Lock control file _before_ reading its content. Fix for a segfault when reading an old control file (bzero() with a negative second argument) storage/maria/ma_control_file.h: changes to the control file module's API storage/maria/ma_init.c: When Maria shuts down cleanly, store max trid into control file. storage/maria/ma_loghandler.c: new prototype storage/maria/ma_recovery.c: During recovery, consult max trid stored in control file, in case it is bigger than what we found in log (case of logs manually removed by user). storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype storage/maria/maria_chk.c: New option --require-control-file (abort if control file not found), --datadir (path for control file (and for logs if --log-dir not specified)), --log-dir (path for logs). Try to open control file when maria_chk starts. storage/maria/maria_read_log.c: new prototype storage/maria/trnman.c: A new function to know max trid in transaction manager storage/maria/trnman_public.h: New function storage/maria/unittest/ma_control_file-t.c: new prototypes. Testing storing and retrieving the max trid to/from control file storage/maria/unittest/ma_test_loghandler-t.c: new prototype storage/maria/unittest/ma_test_loghandler_first_lsn-t.c: new prototype storage/maria/unittest/ma_test_loghandler_max_lsn-t.c: new prototype storage/maria/unittest/ma_test_loghandler_multigroup-t.c: new prototype storage/maria/unittest/ma_test_loghandler_multithread-t.c: new prototype storage/maria/unittest/ma_test_loghandler_noflush-t.c: new prototype storage/maria/unittest/ma_test_loghandler_nologs-t.c: new prototype storage/maria/unittest/ma_test_loghandler_pagecache-t.c: new prototype storage/maria/unittest/ma_test_loghandler_purge-t.c: new prototype
2008-04-04 19:10:53 +02:00
if (unlikely(ma_control_file_write_and_force(lsn, last_logno,
max_trid_in_control_file)))
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
{
translog_unlock();
goto err;
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
translog_unlock();
}
/*
Note that we should not alter memory structures until we have successfully
written the checkpoint record and control file.
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* checkpoint succeeded */
ptr= record_pieces[3].str;
pages_to_flush_before_next_checkpoint= uint4korr(ptr);
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
DBUG_PRINT("checkpoint",("%u pages to flush before next checkpoint",
(uint)pages_to_flush_before_next_checkpoint));
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* compute log's low-water mark */
{
TRANSLOG_ADDRESS log_low_water_mark= min_page_rec_lsn;
set_if_smaller(log_low_water_mark, min_trn_rec_lsn);
set_if_smaller(log_low_water_mark, min_first_undo_lsn);
set_if_smaller(log_low_water_mark, checkpoint_start_log_horizon);
/**
Now purge unneeded logs.
As some systems have an unreliable fsync (drive lying), we could try to
be robust against that: remember a few previous checkpoints in the
control file, and not purge logs immediately... Think about it.
*/
if (translog_purge(log_low_water_mark))
ma_message_no_user(0, "log purging failed");
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
goto end;
err:
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
error= 1;
WL#3071 Maria checkpoint, WL#3072 Maria recovery instead of fprintf(stderr) when a task (with no user connected) gets an error, use my_printf_error(). Flags ME_JUST_WARNING and ME_JUST_INFO added to my_error()/my_printf_error(), which pass it to my_message_sql() which is modified to call the appropriate sql_print_*(). This way recovery can signal its start and end with [Note] and not [ERROR] (but failure with [ERROR]). Recovery's detailed progress (percents etc) still uses stderr as they have to stay on one single line. sql_print_error() changed to use my_progname_short (nicer display). mysql-test-run.pl --gdb/--ddd does not run mysqld, because a breakpoint in mysql_parse is too late to debug startup problems; instead, dev should set the breakpoints it wants and then "run" ("r"). include/my_sys.h: new flags to tell error_handler_hook that this is not an error but an information or warning mysql-test/mysql-test-run.pl: when running with --gdb/--ddd to debug mysqld, breaking at mysql_parse is too late to debug startup problems; now, it does not run mysqld, does not set breakpoints, developer can set as early breakpoints as it wants and is responsible for typing "run" (or "r") mysys/my_init.c: set my_progname_short mysys/my_static.c: my_progname_short added sql/mysqld.cc: * my_message_sql() can now receive info or warning, not only error; this allows mysys to tell the user (or the error log if no user) about an info or warning. Used from Maria. * plugins (or engines like Maria) may want to call my_error(), so set up the error handler hook (my_message_sql) before initializing plugins; otherwise they get my_message_no_curses which is less integrated into mysqld (is just fputs()) * using my_progname_short instead of my_progname, in my_message_sql() (less space on screen) storage/maria/ma_checkpoint.c: fprintf(stderr) -> ma_message_no_user() storage/maria/ma_checkpoint.h: function for any Maria task, not connected to a user (example: checkpoint, recovery; soon could be deleted records purger) to report a message (calls my_printf_error() which, when inside ha_maria, leads to sql_print_*(), and when outside, leads to my_message_no_curses i.e. stderr). storage/maria/ma_recovery.c: To tell that recovery starts and ends we use ma_message_no_user() (sql_print_*() in practice). Detailed progress info still uses stderr as sql_print() cannot put several messages on one line. 071116 18:42:16 [Note] mysqld: Maria engine: starting recovery recovered pages: 0% 67% 100% (0.0 seconds); transactions to roll back: 1 0 (0.0 seconds); tables to flush: 1 0 (0.0 seconds); 071116 18:42:16 [Note] mysqld: Maria engine: recovery done storage/maria/maria_chk.c: my_progname_short moved to mysys storage/maria/maria_read_log.c: my_progname_short moved to mysys storage/myisam/myisamchk.c: my_progname_short moved to mysys
2007-11-16 17:09:51 +01:00
ma_message_no_user(0, "checkpoint failed");
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* we were possibly not able to determine what pages to flush */
pages_to_flush_before_next_checkpoint= 0;
end:
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
for (i= 0; i < (sizeof(record_pieces)/sizeof(record_pieces[0])); i++)
my_free(record_pieces[i].str, MYF(MY_ALLOW_ZERO_PTR));
pthread_mutex_lock(&LOCK_checkpoint);
checkpoint_in_progress= CHECKPOINT_NONE;
checkpoints_total++;
checkpoints_ok_total+= !error;
pthread_mutex_unlock(&LOCK_checkpoint);
DBUG_RETURN(error);
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/**
@brief Initializes the checkpoint module
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
@param interval If one wants the module to create a
thread which will periodically do
checkpoints, and flush dirty pages, in the
background, it should specify a non-zero
interval in seconds. The thread will then be
created and will take checkpoints separated by
approximately 'interval' second.
@note A checkpoint is taken only if there has been some significant
activity since the previous checkpoint. Between checkpoint N and N+1 the
thread flushes all dirty pages which were already dirty at the time of
checkpoint N.
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
@return Operation status
@retval 0 ok
@retval !=0 error
*/
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
int ma_checkpoint_init(ulong interval)
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
{
pthread_t th;
int res= 0;
DBUG_ENTER("ma_checkpoint_init");
checkpoint_inited= TRUE;
checkpoint_thread_die= 2; /* not yet born == dead */
if (pthread_mutex_init(&LOCK_checkpoint, MY_MUTEX_INIT_SLOW) ||
pthread_cond_init(&COND_checkpoint, 0))
res= 1;
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
else if (interval > 0)
{
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
compile_time_assert(sizeof(void *) >= sizeof(ulong));
if (!(res= pthread_create(&th, NULL, ma_checkpoint_background,
(void *)interval)))
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
checkpoint_thread_die= 0; /* thread lives, will have to be killed */
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
DBUG_RETURN(res);
}
* WL#4137 Maria- Framework for testing recovery in mysql-test-run See test maria-recovery.test for a model; all include scripts have an "API" section at start if they do take parameters from outside. * Fixing bug reported by Jani and Monty (when two REDOs about the same page in one group, see ma_blockrec.c). * Fixing small bugs in recovery mysql-test/include/wait_until_connected_again.inc: be sure to enter the loop (the previous query by the caller may not have failed: it could be query; mysqladmin shutdown; call this script). mysql-test/lib/mtr_process.pl: * Through the "expect" file a test can tell mtr that a server crash is expected. What the file contains is irrelevant. Now if its last line starts with "wait", mtr will wait before restarting (it will wait for the last line to not start with "wait"). This is for tests which need to mangle files under the feet of a dead mysqld. * Remove "expect" file before restarting; otherwise there could be a race condition: tests sees server restarted, does something, writes an "expect" file, and then mtr removes that file, then test kills mysqld, and then mtr will never restart it. storage/maria/ma_blockrec.c: - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - fixing bug in applying of REDO_PURGE_BLOCKS in recovery: page_range sometimes has TAIL_BIT set, need to turn it down to know the real page range. - Both bugs are covered in maria-recovery.test storage/maria/ma_checkpoint.c: Capability to, in debug builds only, do some special operations (flush all bitmap and data pages, flush state, flush log) and crash mysqld, to later test recovery. Driven by some --debug=d, symbols. storage/maria/ma_open.c: debugging info storage/maria/ma_pagecache.c: Now that we can _ma_unpin_all_pages() during the REDO phase to set page's LSN, the assertion needs to be relaxed. storage/maria/ma_recovery.c: - open trace file in append mode (useful when a test triggers several recoveries, we see them all). - fixing wrong error detection, it's possible that during recovery we want to open an already open table. - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - we verify that all log records of a group are about the same table, for debugging. mysql-test/r/maria-recovery.result: result mysql-test/t/maria-recovery-master.opt: crash is expected, core file would take room, stack trace would wake pushbuild up. mysql-test/t/maria-recovery.test: Test of recovery from mysql-test (it is already tested as unit tests in ma_test_recovery) (WL#4137) - test that, if recovery is made to start on an empty table it can replay the effects of committed and uncommitted statements (having only the committed ones in the end result). This should be the first test for someone writing code of new REDOs. - test that, if mysqld is crashed and recovery runs we have only committed statements in the end result. Crashes are done in different ways: flush nothing (so, uncommitted statement is often missing from the log => no rollback to do); flush pagecache (implicitely flushes log (WAL)) and flush log, both causes rollbacks; flush log can also flush state (state.records etc) to test recovery of the state (not tested well now as we repair the index anyway). - test of bug found by Jani and Monty in recovery (two REDO about the same page in one group). mysql-test/include/maria_empty_logs.inc: removes logs, to have a clean sheet for testing recovery. mysql-test/include/maria_make_snapshot.inc: copies a table to another directory, or back, or compares both (comparison is not implemented as physical comparison is impossible if an UNDO phase happened). mysql-test/include/maria_make_snapshot_for_comparison.inc: copies tables to another directory so that they can later serve as a comparison reference (they are the good tables, recovery should produce similar ones). mysql-test/include/maria_make_snapshot_for_feeding_recovery.inc: When we want to force recovery to start on old tables, we prepare old tables with this script: we put them in a spare directory. They are later copied back over mysqltest tables while mysqld is dead. We also need to copy back the control file, otherwise mysqld, in recovery, would start from the latest checkpoint: latest checkpoint plus old tables is not a recovery-possible scenario of course. mysql-test/include/maria_verify_recovery.inc: causes mysqld to crash, restores old tables if requested, lets recovery run, compares resulting tables with reference tables by using CHECKSUM TABLE. We don't do any sanity checks on page's LSN in resulting tables, yet.
2007-11-13 17:12:29 +01:00
#ifndef DBUG_OFF
/**
Function used to test recovery: flush some table pieces and then caller
crashes.
@param what_to_flush 0: current bitmap and all data pages
1: state
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
2: all bitmap pages
* WL#4137 Maria- Framework for testing recovery in mysql-test-run See test maria-recovery.test for a model; all include scripts have an "API" section at start if they do take parameters from outside. * Fixing bug reported by Jani and Monty (when two REDOs about the same page in one group, see ma_blockrec.c). * Fixing small bugs in recovery mysql-test/include/wait_until_connected_again.inc: be sure to enter the loop (the previous query by the caller may not have failed: it could be query; mysqladmin shutdown; call this script). mysql-test/lib/mtr_process.pl: * Through the "expect" file a test can tell mtr that a server crash is expected. What the file contains is irrelevant. Now if its last line starts with "wait", mtr will wait before restarting (it will wait for the last line to not start with "wait"). This is for tests which need to mangle files under the feet of a dead mysqld. * Remove "expect" file before restarting; otherwise there could be a race condition: tests sees server restarted, does something, writes an "expect" file, and then mtr removes that file, then test kills mysqld, and then mtr will never restart it. storage/maria/ma_blockrec.c: - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - fixing bug in applying of REDO_PURGE_BLOCKS in recovery: page_range sometimes has TAIL_BIT set, need to turn it down to know the real page range. - Both bugs are covered in maria-recovery.test storage/maria/ma_checkpoint.c: Capability to, in debug builds only, do some special operations (flush all bitmap and data pages, flush state, flush log) and crash mysqld, to later test recovery. Driven by some --debug=d, symbols. storage/maria/ma_open.c: debugging info storage/maria/ma_pagecache.c: Now that we can _ma_unpin_all_pages() during the REDO phase to set page's LSN, the assertion needs to be relaxed. storage/maria/ma_recovery.c: - open trace file in append mode (useful when a test triggers several recoveries, we see them all). - fixing wrong error detection, it's possible that during recovery we want to open an already open table. - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - we verify that all log records of a group are about the same table, for debugging. mysql-test/r/maria-recovery.result: result mysql-test/t/maria-recovery-master.opt: crash is expected, core file would take room, stack trace would wake pushbuild up. mysql-test/t/maria-recovery.test: Test of recovery from mysql-test (it is already tested as unit tests in ma_test_recovery) (WL#4137) - test that, if recovery is made to start on an empty table it can replay the effects of committed and uncommitted statements (having only the committed ones in the end result). This should be the first test for someone writing code of new REDOs. - test that, if mysqld is crashed and recovery runs we have only committed statements in the end result. Crashes are done in different ways: flush nothing (so, uncommitted statement is often missing from the log => no rollback to do); flush pagecache (implicitely flushes log (WAL)) and flush log, both causes rollbacks; flush log can also flush state (state.records etc) to test recovery of the state (not tested well now as we repair the index anyway). - test of bug found by Jani and Monty in recovery (two REDO about the same page in one group). mysql-test/include/maria_empty_logs.inc: removes logs, to have a clean sheet for testing recovery. mysql-test/include/maria_make_snapshot.inc: copies a table to another directory, or back, or compares both (comparison is not implemented as physical comparison is impossible if an UNDO phase happened). mysql-test/include/maria_make_snapshot_for_comparison.inc: copies tables to another directory so that they can later serve as a comparison reference (they are the good tables, recovery should produce similar ones). mysql-test/include/maria_make_snapshot_for_feeding_recovery.inc: When we want to force recovery to start on old tables, we prepare old tables with this script: we put them in a spare directory. They are later copied back over mysqltest tables while mysqld is dead. We also need to copy back the control file, otherwise mysqld, in recovery, would start from the latest checkpoint: latest checkpoint plus old tables is not a recovery-possible scenario of course. mysql-test/include/maria_verify_recovery.inc: causes mysqld to crash, restores old tables if requested, lets recovery run, compares resulting tables with reference tables by using CHECKSUM TABLE. We don't do any sanity checks on page's LSN in resulting tables, yet.
2007-11-13 17:12:29 +01:00
*/
static void flush_all_tables(int what_to_flush)
{
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
int res= 0;
* WL#4137 Maria- Framework for testing recovery in mysql-test-run See test maria-recovery.test for a model; all include scripts have an "API" section at start if they do take parameters from outside. * Fixing bug reported by Jani and Monty (when two REDOs about the same page in one group, see ma_blockrec.c). * Fixing small bugs in recovery mysql-test/include/wait_until_connected_again.inc: be sure to enter the loop (the previous query by the caller may not have failed: it could be query; mysqladmin shutdown; call this script). mysql-test/lib/mtr_process.pl: * Through the "expect" file a test can tell mtr that a server crash is expected. What the file contains is irrelevant. Now if its last line starts with "wait", mtr will wait before restarting (it will wait for the last line to not start with "wait"). This is for tests which need to mangle files under the feet of a dead mysqld. * Remove "expect" file before restarting; otherwise there could be a race condition: tests sees server restarted, does something, writes an "expect" file, and then mtr removes that file, then test kills mysqld, and then mtr will never restart it. storage/maria/ma_blockrec.c: - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - fixing bug in applying of REDO_PURGE_BLOCKS in recovery: page_range sometimes has TAIL_BIT set, need to turn it down to know the real page range. - Both bugs are covered in maria-recovery.test storage/maria/ma_checkpoint.c: Capability to, in debug builds only, do some special operations (flush all bitmap and data pages, flush state, flush log) and crash mysqld, to later test recovery. Driven by some --debug=d, symbols. storage/maria/ma_open.c: debugging info storage/maria/ma_pagecache.c: Now that we can _ma_unpin_all_pages() during the REDO phase to set page's LSN, the assertion needs to be relaxed. storage/maria/ma_recovery.c: - open trace file in append mode (useful when a test triggers several recoveries, we see them all). - fixing wrong error detection, it's possible that during recovery we want to open an already open table. - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - we verify that all log records of a group are about the same table, for debugging. mysql-test/r/maria-recovery.result: result mysql-test/t/maria-recovery-master.opt: crash is expected, core file would take room, stack trace would wake pushbuild up. mysql-test/t/maria-recovery.test: Test of recovery from mysql-test (it is already tested as unit tests in ma_test_recovery) (WL#4137) - test that, if recovery is made to start on an empty table it can replay the effects of committed and uncommitted statements (having only the committed ones in the end result). This should be the first test for someone writing code of new REDOs. - test that, if mysqld is crashed and recovery runs we have only committed statements in the end result. Crashes are done in different ways: flush nothing (so, uncommitted statement is often missing from the log => no rollback to do); flush pagecache (implicitely flushes log (WAL)) and flush log, both causes rollbacks; flush log can also flush state (state.records etc) to test recovery of the state (not tested well now as we repair the index anyway). - test of bug found by Jani and Monty in recovery (two REDO about the same page in one group). mysql-test/include/maria_empty_logs.inc: removes logs, to have a clean sheet for testing recovery. mysql-test/include/maria_make_snapshot.inc: copies a table to another directory, or back, or compares both (comparison is not implemented as physical comparison is impossible if an UNDO phase happened). mysql-test/include/maria_make_snapshot_for_comparison.inc: copies tables to another directory so that they can later serve as a comparison reference (they are the good tables, recovery should produce similar ones). mysql-test/include/maria_make_snapshot_for_feeding_recovery.inc: When we want to force recovery to start on old tables, we prepare old tables with this script: we put them in a spare directory. They are later copied back over mysqltest tables while mysqld is dead. We also need to copy back the control file, otherwise mysqld, in recovery, would start from the latest checkpoint: latest checkpoint plus old tables is not a recovery-possible scenario of course. mysql-test/include/maria_verify_recovery.inc: causes mysqld to crash, restores old tables if requested, lets recovery run, compares resulting tables with reference tables by using CHECKSUM TABLE. We don't do any sanity checks on page's LSN in resulting tables, yet.
2007-11-13 17:12:29 +01:00
LIST *pos; /**< to iterate over open tables */
pthread_mutex_lock(&THR_LOCK_maria);
for (pos= maria_open_list; pos; pos= pos->next)
{
MARIA_HA *info= (MARIA_HA*)pos->data;
if (info->s->now_transactional)
{
switch (what_to_flush)
{
case 0:
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
res= _ma_flush_table_files(info, MARIA_FLUSH_DATA | MARIA_FLUSH_INDEX,
WL#3072 Maria Recovery All statements doing an implicit commit now also do one in Maria. This is useful because LOCK TABLES; REPAIR; crash; is not rollback-able, the implicit commit of REPAIR avoid that Recovery tries to rollback and fails. Fix for BUG#33827 "COMMIT AND CHAIN causes serious Valgrind error" (maybe not the definite one, depends on the assigned dev). mysql-test/t/maria-recovery.test: test of REPAIR's implicit commit. I cannot commit the result file because maria-recovery fails in vanilla tree (seen in pushbuild) but its new section looks like: repair table t1; Table Op Msg_type Msg_text mysqltest.t1 repair status OK insert into t1 values(2); select * from t1; a 1 2 3 SET SESSION debug="+d,maria_flush_whole_log,maria_flush_whole_page_cache,maria_crash"; * crashing mysqld intentionally set global maria_checkpoint_interval=1; ERROR HY000: Lost connection to MySQL server during query * recovery happens check table t1 extended; Table Op Msg_type Msg_text mysqltest.t1 check status OK * testing that checksum after recovery is as expected Checksum-check failure use mysqltest; select * from t1; a 1 3 Which is as it should be. sql/rpl_injector.cc: fix for BUG#33827 sql/sql_parse.cc: - All DDLs and mysql_admin_table() (REPAIR etc) use end_actrive_trans() to do an implicit commit so we add there an implicit commit of the Maria transaction. - Fix for BUG#33827 storage/maria/ha_maria.cc: - A method to do implicit commit in Maria - After an implicit commit, if it was under LOCK TABLES, the locked tables have a stale file->trn: update it. storage/maria/ha_maria.h: new static method storage/maria/ma_check.c: bugfix: this disabling of transactionality had the effect that if LOCK TABLES; REPAIR; INSERT then the INSERT ran non-transactional (so couldn't be undone in case of crash, if, by bad chance, its effect on pages went to disk). storage/maria/ma_checkpoint.c: indentation storage/maria/ma_recovery.c: dbug statements storage/maria/trnman.c: When doing an implicit commit we need to know the number of locked tables of the committed transaction and copy it to the new transaction storage/maria/trnman_public.h: prototype change
2008-01-11 22:48:54 +01:00
FLUSH_KEEP, FLUSH_KEEP);
* WL#4137 Maria- Framework for testing recovery in mysql-test-run See test maria-recovery.test for a model; all include scripts have an "API" section at start if they do take parameters from outside. * Fixing bug reported by Jani and Monty (when two REDOs about the same page in one group, see ma_blockrec.c). * Fixing small bugs in recovery mysql-test/include/wait_until_connected_again.inc: be sure to enter the loop (the previous query by the caller may not have failed: it could be query; mysqladmin shutdown; call this script). mysql-test/lib/mtr_process.pl: * Through the "expect" file a test can tell mtr that a server crash is expected. What the file contains is irrelevant. Now if its last line starts with "wait", mtr will wait before restarting (it will wait for the last line to not start with "wait"). This is for tests which need to mangle files under the feet of a dead mysqld. * Remove "expect" file before restarting; otherwise there could be a race condition: tests sees server restarted, does something, writes an "expect" file, and then mtr removes that file, then test kills mysqld, and then mtr will never restart it. storage/maria/ma_blockrec.c: - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - fixing bug in applying of REDO_PURGE_BLOCKS in recovery: page_range sometimes has TAIL_BIT set, need to turn it down to know the real page range. - Both bugs are covered in maria-recovery.test storage/maria/ma_checkpoint.c: Capability to, in debug builds only, do some special operations (flush all bitmap and data pages, flush state, flush log) and crash mysqld, to later test recovery. Driven by some --debug=d, symbols. storage/maria/ma_open.c: debugging info storage/maria/ma_pagecache.c: Now that we can _ma_unpin_all_pages() during the REDO phase to set page's LSN, the assertion needs to be relaxed. storage/maria/ma_recovery.c: - open trace file in append mode (useful when a test triggers several recoveries, we see them all). - fixing wrong error detection, it's possible that during recovery we want to open an already open table. - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - we verify that all log records of a group are about the same table, for debugging. mysql-test/r/maria-recovery.result: result mysql-test/t/maria-recovery-master.opt: crash is expected, core file would take room, stack trace would wake pushbuild up. mysql-test/t/maria-recovery.test: Test of recovery from mysql-test (it is already tested as unit tests in ma_test_recovery) (WL#4137) - test that, if recovery is made to start on an empty table it can replay the effects of committed and uncommitted statements (having only the committed ones in the end result). This should be the first test for someone writing code of new REDOs. - test that, if mysqld is crashed and recovery runs we have only committed statements in the end result. Crashes are done in different ways: flush nothing (so, uncommitted statement is often missing from the log => no rollback to do); flush pagecache (implicitely flushes log (WAL)) and flush log, both causes rollbacks; flush log can also flush state (state.records etc) to test recovery of the state (not tested well now as we repair the index anyway). - test of bug found by Jani and Monty in recovery (two REDO about the same page in one group). mysql-test/include/maria_empty_logs.inc: removes logs, to have a clean sheet for testing recovery. mysql-test/include/maria_make_snapshot.inc: copies a table to another directory, or back, or compares both (comparison is not implemented as physical comparison is impossible if an UNDO phase happened). mysql-test/include/maria_make_snapshot_for_comparison.inc: copies tables to another directory so that they can later serve as a comparison reference (they are the good tables, recovery should produce similar ones). mysql-test/include/maria_make_snapshot_for_feeding_recovery.inc: When we want to force recovery to start on old tables, we prepare old tables with this script: we put them in a spare directory. They are later copied back over mysqltest tables while mysqld is dead. We also need to copy back the control file, otherwise mysqld, in recovery, would start from the latest checkpoint: latest checkpoint plus old tables is not a recovery-possible scenario of course. mysql-test/include/maria_verify_recovery.inc: causes mysqld to crash, restores old tables if requested, lets recovery run, compares resulting tables with reference tables by using CHECKSUM TABLE. We don't do any sanity checks on page's LSN in resulting tables, yet.
2007-11-13 17:12:29 +01:00
break;
case 1:
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
res= _ma_state_info_write(info->s, 1|4);
* WL#4137 Maria- Framework for testing recovery in mysql-test-run See test maria-recovery.test for a model; all include scripts have an "API" section at start if they do take parameters from outside. * Fixing bug reported by Jani and Monty (when two REDOs about the same page in one group, see ma_blockrec.c). * Fixing small bugs in recovery mysql-test/include/wait_until_connected_again.inc: be sure to enter the loop (the previous query by the caller may not have failed: it could be query; mysqladmin shutdown; call this script). mysql-test/lib/mtr_process.pl: * Through the "expect" file a test can tell mtr that a server crash is expected. What the file contains is irrelevant. Now if its last line starts with "wait", mtr will wait before restarting (it will wait for the last line to not start with "wait"). This is for tests which need to mangle files under the feet of a dead mysqld. * Remove "expect" file before restarting; otherwise there could be a race condition: tests sees server restarted, does something, writes an "expect" file, and then mtr removes that file, then test kills mysqld, and then mtr will never restart it. storage/maria/ma_blockrec.c: - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - fixing bug in applying of REDO_PURGE_BLOCKS in recovery: page_range sometimes has TAIL_BIT set, need to turn it down to know the real page range. - Both bugs are covered in maria-recovery.test storage/maria/ma_checkpoint.c: Capability to, in debug builds only, do some special operations (flush all bitmap and data pages, flush state, flush log) and crash mysqld, to later test recovery. Driven by some --debug=d, symbols. storage/maria/ma_open.c: debugging info storage/maria/ma_pagecache.c: Now that we can _ma_unpin_all_pages() during the REDO phase to set page's LSN, the assertion needs to be relaxed. storage/maria/ma_recovery.c: - open trace file in append mode (useful when a test triggers several recoveries, we see them all). - fixing wrong error detection, it's possible that during recovery we want to open an already open table. - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - we verify that all log records of a group are about the same table, for debugging. mysql-test/r/maria-recovery.result: result mysql-test/t/maria-recovery-master.opt: crash is expected, core file would take room, stack trace would wake pushbuild up. mysql-test/t/maria-recovery.test: Test of recovery from mysql-test (it is already tested as unit tests in ma_test_recovery) (WL#4137) - test that, if recovery is made to start on an empty table it can replay the effects of committed and uncommitted statements (having only the committed ones in the end result). This should be the first test for someone writing code of new REDOs. - test that, if mysqld is crashed and recovery runs we have only committed statements in the end result. Crashes are done in different ways: flush nothing (so, uncommitted statement is often missing from the log => no rollback to do); flush pagecache (implicitely flushes log (WAL)) and flush log, both causes rollbacks; flush log can also flush state (state.records etc) to test recovery of the state (not tested well now as we repair the index anyway). - test of bug found by Jani and Monty in recovery (two REDO about the same page in one group). mysql-test/include/maria_empty_logs.inc: removes logs, to have a clean sheet for testing recovery. mysql-test/include/maria_make_snapshot.inc: copies a table to another directory, or back, or compares both (comparison is not implemented as physical comparison is impossible if an UNDO phase happened). mysql-test/include/maria_make_snapshot_for_comparison.inc: copies tables to another directory so that they can later serve as a comparison reference (they are the good tables, recovery should produce similar ones). mysql-test/include/maria_make_snapshot_for_feeding_recovery.inc: When we want to force recovery to start on old tables, we prepare old tables with this script: we put them in a spare directory. They are later copied back over mysqltest tables while mysqld is dead. We also need to copy back the control file, otherwise mysqld, in recovery, would start from the latest checkpoint: latest checkpoint plus old tables is not a recovery-possible scenario of course. mysql-test/include/maria_verify_recovery.inc: causes mysqld to crash, restores old tables if requested, lets recovery run, compares resulting tables with reference tables by using CHECKSUM TABLE. We don't do any sanity checks on page's LSN in resulting tables, yet.
2007-11-13 17:12:29 +01:00
DBUG_PRINT("maria_flush_states",
("is_of_horizon: LSN (%lu,0x%lx)",
LSN_IN_PARTS(info->s->state.is_of_horizon)));
break;
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
case 2:
res= _ma_bitmap_flush_all(info->s);
break;
* WL#4137 Maria- Framework for testing recovery in mysql-test-run See test maria-recovery.test for a model; all include scripts have an "API" section at start if they do take parameters from outside. * Fixing bug reported by Jani and Monty (when two REDOs about the same page in one group, see ma_blockrec.c). * Fixing small bugs in recovery mysql-test/include/wait_until_connected_again.inc: be sure to enter the loop (the previous query by the caller may not have failed: it could be query; mysqladmin shutdown; call this script). mysql-test/lib/mtr_process.pl: * Through the "expect" file a test can tell mtr that a server crash is expected. What the file contains is irrelevant. Now if its last line starts with "wait", mtr will wait before restarting (it will wait for the last line to not start with "wait"). This is for tests which need to mangle files under the feet of a dead mysqld. * Remove "expect" file before restarting; otherwise there could be a race condition: tests sees server restarted, does something, writes an "expect" file, and then mtr removes that file, then test kills mysqld, and then mtr will never restart it. storage/maria/ma_blockrec.c: - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - fixing bug in applying of REDO_PURGE_BLOCKS in recovery: page_range sometimes has TAIL_BIT set, need to turn it down to know the real page range. - Both bugs are covered in maria-recovery.test storage/maria/ma_checkpoint.c: Capability to, in debug builds only, do some special operations (flush all bitmap and data pages, flush state, flush log) and crash mysqld, to later test recovery. Driven by some --debug=d, symbols. storage/maria/ma_open.c: debugging info storage/maria/ma_pagecache.c: Now that we can _ma_unpin_all_pages() during the REDO phase to set page's LSN, the assertion needs to be relaxed. storage/maria/ma_recovery.c: - open trace file in append mode (useful when a test triggers several recoveries, we see them all). - fixing wrong error detection, it's possible that during recovery we want to open an already open table. - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - we verify that all log records of a group are about the same table, for debugging. mysql-test/r/maria-recovery.result: result mysql-test/t/maria-recovery-master.opt: crash is expected, core file would take room, stack trace would wake pushbuild up. mysql-test/t/maria-recovery.test: Test of recovery from mysql-test (it is already tested as unit tests in ma_test_recovery) (WL#4137) - test that, if recovery is made to start on an empty table it can replay the effects of committed and uncommitted statements (having only the committed ones in the end result). This should be the first test for someone writing code of new REDOs. - test that, if mysqld is crashed and recovery runs we have only committed statements in the end result. Crashes are done in different ways: flush nothing (so, uncommitted statement is often missing from the log => no rollback to do); flush pagecache (implicitely flushes log (WAL)) and flush log, both causes rollbacks; flush log can also flush state (state.records etc) to test recovery of the state (not tested well now as we repair the index anyway). - test of bug found by Jani and Monty in recovery (two REDO about the same page in one group). mysql-test/include/maria_empty_logs.inc: removes logs, to have a clean sheet for testing recovery. mysql-test/include/maria_make_snapshot.inc: copies a table to another directory, or back, or compares both (comparison is not implemented as physical comparison is impossible if an UNDO phase happened). mysql-test/include/maria_make_snapshot_for_comparison.inc: copies tables to another directory so that they can later serve as a comparison reference (they are the good tables, recovery should produce similar ones). mysql-test/include/maria_make_snapshot_for_feeding_recovery.inc: When we want to force recovery to start on old tables, we prepare old tables with this script: we put them in a spare directory. They are later copied back over mysqltest tables while mysqld is dead. We also need to copy back the control file, otherwise mysqld, in recovery, would start from the latest checkpoint: latest checkpoint plus old tables is not a recovery-possible scenario of course. mysql-test/include/maria_verify_recovery.inc: causes mysqld to crash, restores old tables if requested, lets recovery run, compares resulting tables with reference tables by using CHECKSUM TABLE. We don't do any sanity checks on page's LSN in resulting tables, yet.
2007-11-13 17:12:29 +01:00
}
}
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
DBUG_ASSERT(res == 0);
* WL#4137 Maria- Framework for testing recovery in mysql-test-run See test maria-recovery.test for a model; all include scripts have an "API" section at start if they do take parameters from outside. * Fixing bug reported by Jani and Monty (when two REDOs about the same page in one group, see ma_blockrec.c). * Fixing small bugs in recovery mysql-test/include/wait_until_connected_again.inc: be sure to enter the loop (the previous query by the caller may not have failed: it could be query; mysqladmin shutdown; call this script). mysql-test/lib/mtr_process.pl: * Through the "expect" file a test can tell mtr that a server crash is expected. What the file contains is irrelevant. Now if its last line starts with "wait", mtr will wait before restarting (it will wait for the last line to not start with "wait"). This is for tests which need to mangle files under the feet of a dead mysqld. * Remove "expect" file before restarting; otherwise there could be a race condition: tests sees server restarted, does something, writes an "expect" file, and then mtr removes that file, then test kills mysqld, and then mtr will never restart it. storage/maria/ma_blockrec.c: - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - fixing bug in applying of REDO_PURGE_BLOCKS in recovery: page_range sometimes has TAIL_BIT set, need to turn it down to know the real page range. - Both bugs are covered in maria-recovery.test storage/maria/ma_checkpoint.c: Capability to, in debug builds only, do some special operations (flush all bitmap and data pages, flush state, flush log) and crash mysqld, to later test recovery. Driven by some --debug=d, symbols. storage/maria/ma_open.c: debugging info storage/maria/ma_pagecache.c: Now that we can _ma_unpin_all_pages() during the REDO phase to set page's LSN, the assertion needs to be relaxed. storage/maria/ma_recovery.c: - open trace file in append mode (useful when a test triggers several recoveries, we see them all). - fixing wrong error detection, it's possible that during recovery we want to open an already open table. - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - we verify that all log records of a group are about the same table, for debugging. mysql-test/r/maria-recovery.result: result mysql-test/t/maria-recovery-master.opt: crash is expected, core file would take room, stack trace would wake pushbuild up. mysql-test/t/maria-recovery.test: Test of recovery from mysql-test (it is already tested as unit tests in ma_test_recovery) (WL#4137) - test that, if recovery is made to start on an empty table it can replay the effects of committed and uncommitted statements (having only the committed ones in the end result). This should be the first test for someone writing code of new REDOs. - test that, if mysqld is crashed and recovery runs we have only committed statements in the end result. Crashes are done in different ways: flush nothing (so, uncommitted statement is often missing from the log => no rollback to do); flush pagecache (implicitely flushes log (WAL)) and flush log, both causes rollbacks; flush log can also flush state (state.records etc) to test recovery of the state (not tested well now as we repair the index anyway). - test of bug found by Jani and Monty in recovery (two REDO about the same page in one group). mysql-test/include/maria_empty_logs.inc: removes logs, to have a clean sheet for testing recovery. mysql-test/include/maria_make_snapshot.inc: copies a table to another directory, or back, or compares both (comparison is not implemented as physical comparison is impossible if an UNDO phase happened). mysql-test/include/maria_make_snapshot_for_comparison.inc: copies tables to another directory so that they can later serve as a comparison reference (they are the good tables, recovery should produce similar ones). mysql-test/include/maria_make_snapshot_for_feeding_recovery.inc: When we want to force recovery to start on old tables, we prepare old tables with this script: we put them in a spare directory. They are later copied back over mysqltest tables while mysqld is dead. We also need to copy back the control file, otherwise mysqld, in recovery, would start from the latest checkpoint: latest checkpoint plus old tables is not a recovery-possible scenario of course. mysql-test/include/maria_verify_recovery.inc: causes mysqld to crash, restores old tables if requested, lets recovery run, compares resulting tables with reference tables by using CHECKSUM TABLE. We don't do any sanity checks on page's LSN in resulting tables, yet.
2007-11-13 17:12:29 +01:00
}
pthread_mutex_unlock(&THR_LOCK_maria);
}
#endif
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/**
@brief Destroys the checkpoint module
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
void ma_checkpoint_end(void)
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
DBUG_ENTER("ma_checkpoint_end");
WL#3072 - Maria Recovery * to honour WAL we now force the whole log when flushing a bitmap page. * ability to intentionally crash in various places for recovery testing * bugfix (dirty pages list found in checkpoint record was ignored) * smaller checkpoint record * misc small cleanups and comments mysql-test/include/maria_empty_logs.inc: maria-purge.test creates ~11 logs, remove them all mysql-test/r/maria-recovery-bitmap.result: result is good; without the _ma_bitmap_get_log_address() call, we got check error Bitmap at 0 has pages reserved outside of data file length mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery-bitmap.test: enable test of "bitmap-flush should flush whole log otherwise corrupted data file (bitmap ahead of data pages)". mysql-test/t/maria-recovery.test: test of checkpoint sql/sql_table.cc: comment storage/maria/ha_maria.cc: _ma_reenable_logging_for_table() now includes file->trn=0. At the end of repair() we don't need to re-enable logging, it is done already by caller (like copy_data_between_tables()); it sounds strange that this function could decide to re-enable, it should be up to caller who knows what other operations it plans. Removing this line led to assertion failure in maria_lock_database(F_UNLCK), fixed by removing the assertion: maria_lock_database() is here called in a context where F_UNLCK does not make the table visible to others so assertion is excessive, and external_lock() is already designed to honour the asserted condition. Ability to crash at the end of bulk insert when indices have been enabled. storage/maria/ma_bitmap.c: Better use pagecache_file_init() than set pagecache callbacks directly; and a new function to set those callbacks for bitmap so that we can reuse it. _ma_bitmap_get_log_address() is a pagecache get_log_address callback which causes the whole log to be flushed when a bitmap page is flushed by the page cache. This was required by WAL. storage/maria/ma_blockrec.c: get_log_address pagecache callback for data (non bitmap) pages: just reads the LSN from the page's content, like was hard-coded before in ma_pagecache.c. storage/maria/ma_blockrec.h: functions which need to be exported storage/maria/ma_check.c: create_new_data_handle() can be static. Ability to crash after rebuilding the index in OPTIMIZE, in REPAIR. my_lock() implemented already. storage/maria/ma_checkpoint.c: As MARIA_SHARE* is now accessible to pagecache_collect_changed_blocks_LSN(), we don't need to store kfile/dfile descriptors in checkpoint record, 2-byte-id of the table plus one byte to say if this is data or index file is enough. So we go from 4+4 bytes per table down to 2+1. storage/maria/ma_commit.c: removing duplicate functions (see _ma_tmp_disable_logging_for_table()) storage/maria/ma_extra.c: Monty fixed storage/maria/ma_key_recover.c: comment storage/maria/ma_locking.c: Sometimes other code does funny things with maria_lock_database(), like ha_maria::repair() calling it at start and end without going through ha_maria::external_lock(). So it happens that maria_lock_database() is called with now_transactional!=born_transactional. storage/maria/ma_loghandler.c: update to new prototype storage/maria/ma_open.c: set_data|index_pagecache_callbacks() need to be exported as they are now called when disabling/enabling transactionality. storage/maria/ma_pagecache.c: Removing PAGE_LSN_OFFSET, as much of the code relies on it being 0 anyway (let's not give impression we can just change this constant). When flushing a page to disk, call the get_log_address callback to know up to which LSN the log should be flushed. As we now can access MARIA_SHARE* we can know share->id and store it into the checkpoint record; we thus go from 4 bytes per dirty page to 2+1. storage/maria/ma_pagecache.h: get_log_address callback storage/maria/ma_panic.c: No reason to reset pagecache callbacks in HA_PANIC_READ: all we do is reopen files if they were closed; callbacks should be in place already as 'info' exists; we just want to modify the file descriptors, not the full PAGECACHE_FILE structure. If we open data file and it was closed, share->bitmap.file needs to be set. Note that the modified code is disabled anyway. storage/maria/ma_recovery.c: Checkpoint record does not contain kfile/dfile descriptors anymore so code can be simplified. Hash key in all_dirty_pages is not made from file_descriptor & pageno anymore, but index_or_data & table-short-id & pageno. If a table's create_rename_lsn is higher than record's LSN, we skip the table and don't fail if it's corrupted (because the LSNs say that we don't have to look at this table). If a table is skipped (for example due to create_rename_lsn), its UNDOs still cause undo_lsn to advance; this is so that if later we notice the transaction has to rollback we fail (as table should not be skipped in this case). Fixing a bug: the dirty_pages list was never used, because the LSN below which it was used was the minimum rec_lsn of dirty pages! It is now the min(checkpoint_start_log_horizon, min(trn's rec_lsn)). When we disable/reenable transactionality, we modify pagecache callbacks (needed for example for get_log_address: changing share->page_type is not enough anymore). storage/maria/ma_write.c: 'records' and 'checksum' are protected: they are updated under log's mutex in write-hooks when UNDO is written. storage/maria/maria_chk.c: remove use of duplicate functions. storage/maria/maria_def.h: set_data|index_pagecache_callbacks() need to be exported; _ma_reenable_logging_for_table() changes to a real function. storage/maria/unittest/ma_pagecache_consist.c: new prototype storage/maria/unittest/ma_pagecache_single.c: new prototype storage/maria/unittest/ma_test_loghandler_pagecache-t.c: new prototype
2007-12-30 21:32:07 +01:00
/*
Some intentional crash methods, usually triggered by
SET MARIA_CHECKPOINT_INTERVAL=X
*/
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
DBUG_EXECUTE_IF("maria_flush_bitmap",
{
DBUG_PRINT("maria_flush_bitmap", ("now"));
flush_all_tables(2);
});
* WL#4137 Maria- Framework for testing recovery in mysql-test-run See test maria-recovery.test for a model; all include scripts have an "API" section at start if they do take parameters from outside. * Fixing bug reported by Jani and Monty (when two REDOs about the same page in one group, see ma_blockrec.c). * Fixing small bugs in recovery mysql-test/include/wait_until_connected_again.inc: be sure to enter the loop (the previous query by the caller may not have failed: it could be query; mysqladmin shutdown; call this script). mysql-test/lib/mtr_process.pl: * Through the "expect" file a test can tell mtr that a server crash is expected. What the file contains is irrelevant. Now if its last line starts with "wait", mtr will wait before restarting (it will wait for the last line to not start with "wait"). This is for tests which need to mangle files under the feet of a dead mysqld. * Remove "expect" file before restarting; otherwise there could be a race condition: tests sees server restarted, does something, writes an "expect" file, and then mtr removes that file, then test kills mysqld, and then mtr will never restart it. storage/maria/ma_blockrec.c: - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - fixing bug in applying of REDO_PURGE_BLOCKS in recovery: page_range sometimes has TAIL_BIT set, need to turn it down to know the real page range. - Both bugs are covered in maria-recovery.test storage/maria/ma_checkpoint.c: Capability to, in debug builds only, do some special operations (flush all bitmap and data pages, flush state, flush log) and crash mysqld, to later test recovery. Driven by some --debug=d, symbols. storage/maria/ma_open.c: debugging info storage/maria/ma_pagecache.c: Now that we can _ma_unpin_all_pages() during the REDO phase to set page's LSN, the assertion needs to be relaxed. storage/maria/ma_recovery.c: - open trace file in append mode (useful when a test triggers several recoveries, we see them all). - fixing wrong error detection, it's possible that during recovery we want to open an already open table. - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - we verify that all log records of a group are about the same table, for debugging. mysql-test/r/maria-recovery.result: result mysql-test/t/maria-recovery-master.opt: crash is expected, core file would take room, stack trace would wake pushbuild up. mysql-test/t/maria-recovery.test: Test of recovery from mysql-test (it is already tested as unit tests in ma_test_recovery) (WL#4137) - test that, if recovery is made to start on an empty table it can replay the effects of committed and uncommitted statements (having only the committed ones in the end result). This should be the first test for someone writing code of new REDOs. - test that, if mysqld is crashed and recovery runs we have only committed statements in the end result. Crashes are done in different ways: flush nothing (so, uncommitted statement is often missing from the log => no rollback to do); flush pagecache (implicitely flushes log (WAL)) and flush log, both causes rollbacks; flush log can also flush state (state.records etc) to test recovery of the state (not tested well now as we repair the index anyway). - test of bug found by Jani and Monty in recovery (two REDO about the same page in one group). mysql-test/include/maria_empty_logs.inc: removes logs, to have a clean sheet for testing recovery. mysql-test/include/maria_make_snapshot.inc: copies a table to another directory, or back, or compares both (comparison is not implemented as physical comparison is impossible if an UNDO phase happened). mysql-test/include/maria_make_snapshot_for_comparison.inc: copies tables to another directory so that they can later serve as a comparison reference (they are the good tables, recovery should produce similar ones). mysql-test/include/maria_make_snapshot_for_feeding_recovery.inc: When we want to force recovery to start on old tables, we prepare old tables with this script: we put them in a spare directory. They are later copied back over mysqltest tables while mysqld is dead. We also need to copy back the control file, otherwise mysqld, in recovery, would start from the latest checkpoint: latest checkpoint plus old tables is not a recovery-possible scenario of course. mysql-test/include/maria_verify_recovery.inc: causes mysqld to crash, restores old tables if requested, lets recovery run, compares resulting tables with reference tables by using CHECKSUM TABLE. We don't do any sanity checks on page's LSN in resulting tables, yet.
2007-11-13 17:12:29 +01:00
DBUG_EXECUTE_IF("maria_flush_whole_page_cache",
{
DBUG_PRINT("maria_flush_whole_page_cache", ("now"));
flush_all_tables(0);
});
DBUG_EXECUTE_IF("maria_flush_whole_log",
{
DBUG_PRINT("maria_flush_whole_log", ("now"));
translog_flush(translog_get_horizon());
});
/*
Note that for WAL reasons, maria_flush_states requires
maria_flush_whole_log.
*/
DBUG_EXECUTE_IF("maria_flush_states",
{
DBUG_PRINT("maria_flush_states", ("now"));
flush_all_tables(1);
});
DBUG_EXECUTE_IF("maria_crash",
{ DBUG_PRINT("maria_crash", ("now")); DBUG_ABORT(); });
* WL#4137 Maria- Framework for testing recovery in mysql-test-run See test maria-recovery.test for a model; all include scripts have an "API" section at start if they do take parameters from outside. * Fixing bug reported by Jani and Monty (when two REDOs about the same page in one group, see ma_blockrec.c). * Fixing small bugs in recovery mysql-test/include/wait_until_connected_again.inc: be sure to enter the loop (the previous query by the caller may not have failed: it could be query; mysqladmin shutdown; call this script). mysql-test/lib/mtr_process.pl: * Through the "expect" file a test can tell mtr that a server crash is expected. What the file contains is irrelevant. Now if its last line starts with "wait", mtr will wait before restarting (it will wait for the last line to not start with "wait"). This is for tests which need to mangle files under the feet of a dead mysqld. * Remove "expect" file before restarting; otherwise there could be a race condition: tests sees server restarted, does something, writes an "expect" file, and then mtr removes that file, then test kills mysqld, and then mtr will never restart it. storage/maria/ma_blockrec.c: - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - fixing bug in applying of REDO_PURGE_BLOCKS in recovery: page_range sometimes has TAIL_BIT set, need to turn it down to know the real page range. - Both bugs are covered in maria-recovery.test storage/maria/ma_checkpoint.c: Capability to, in debug builds only, do some special operations (flush all bitmap and data pages, flush state, flush log) and crash mysqld, to later test recovery. Driven by some --debug=d, symbols. storage/maria/ma_open.c: debugging info storage/maria/ma_pagecache.c: Now that we can _ma_unpin_all_pages() during the REDO phase to set page's LSN, the assertion needs to be relaxed. storage/maria/ma_recovery.c: - open trace file in append mode (useful when a test triggers several recoveries, we see them all). - fixing wrong error detection, it's possible that during recovery we want to open an already open table. - when applying a REDO in recovery, we don't anymore put UNDO's LSN on the page at once; indeed if in this REDO's group there comes another REDO for the same page it would be wrongly skipped. Instead, we keep pages pinned, don't change their LSN. When done with all REDOs of the group we unpin them and stamp them with UNDO's LSN. - we verify that all log records of a group are about the same table, for debugging. mysql-test/r/maria-recovery.result: result mysql-test/t/maria-recovery-master.opt: crash is expected, core file would take room, stack trace would wake pushbuild up. mysql-test/t/maria-recovery.test: Test of recovery from mysql-test (it is already tested as unit tests in ma_test_recovery) (WL#4137) - test that, if recovery is made to start on an empty table it can replay the effects of committed and uncommitted statements (having only the committed ones in the end result). This should be the first test for someone writing code of new REDOs. - test that, if mysqld is crashed and recovery runs we have only committed statements in the end result. Crashes are done in different ways: flush nothing (so, uncommitted statement is often missing from the log => no rollback to do); flush pagecache (implicitely flushes log (WAL)) and flush log, both causes rollbacks; flush log can also flush state (state.records etc) to test recovery of the state (not tested well now as we repair the index anyway). - test of bug found by Jani and Monty in recovery (two REDO about the same page in one group). mysql-test/include/maria_empty_logs.inc: removes logs, to have a clean sheet for testing recovery. mysql-test/include/maria_make_snapshot.inc: copies a table to another directory, or back, or compares both (comparison is not implemented as physical comparison is impossible if an UNDO phase happened). mysql-test/include/maria_make_snapshot_for_comparison.inc: copies tables to another directory so that they can later serve as a comparison reference (they are the good tables, recovery should produce similar ones). mysql-test/include/maria_make_snapshot_for_feeding_recovery.inc: When we want to force recovery to start on old tables, we prepare old tables with this script: we put them in a spare directory. They are later copied back over mysqltest tables while mysqld is dead. We also need to copy back the control file, otherwise mysqld, in recovery, would start from the latest checkpoint: latest checkpoint plus old tables is not a recovery-possible scenario of course. mysql-test/include/maria_verify_recovery.inc: causes mysqld to crash, restores old tables if requested, lets recovery run, compares resulting tables with reference tables by using CHECKSUM TABLE. We don't do any sanity checks on page's LSN in resulting tables, yet.
2007-11-13 17:12:29 +01:00
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
if (checkpoint_inited)
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pthread_mutex_lock(&LOCK_checkpoint);
if (checkpoint_thread_die != 2) /* thread was started ok */
{
DBUG_PRINT("info",("killing Maria background checkpoint thread"));
checkpoint_thread_die= 1; /* kill it */
do /* and wait for it to be dead */
{
/* wake it up if it was in a sleep */
pthread_cond_broadcast(&COND_checkpoint);
DBUG_PRINT("info",("waiting for Maria background checkpoint thread"
" to die"));
pthread_cond_wait(&COND_checkpoint, &LOCK_checkpoint);
}
while (checkpoint_thread_die != 2);
}
pthread_mutex_unlock(&LOCK_checkpoint);
my_free((uchar *)dfiles, MYF(MY_ALLOW_ZERO_PTR));
my_free((uchar *)kfiles, MYF(MY_ALLOW_ZERO_PTR));
WL#3072 Maria recovery Misc changes: - fix for benign Valgrind error, compiler warnings - fix for a segfault in execution of maria_delete_all_rows() and one when taking multiple checkpoints - fix for too paranoid assertion - adding ability to take checkpoints at the end of the REDO phase and at the end of recovery. - other minor changes storage/maria/ha_maria.cc: The checkpoint done after Recovery is finished, is moved to maria_recover(). storage/maria/ma_bitmap.c: fix for Valgrind error: the "shadow debug copy" of the bitmap page started unitialized and so ma_print_bitmap() would use it uninitialized storage/maria/ma_checkpoint.c: * reset pointers to NULL after freeing them, or we segfault at next checkpoint in my_realloc(). * fix for compiler warnings. storage/maria/ma_delete_all.c: info->trn is NULL for non-transactional tables storage/maria/ma_locking.c: correct assertion (it fired wrongly in execution of REDO_DROP_TABLE due to the maria_extra(HA_PREPARE_FOR_DROP)->_ma_decrement_open_count() ->maria_lock_database(F_UNLCK); another solution would have been to not call _ma_decrement_open_count() (it's ok to have a wrong open count in a table which we are dropping), but the same problem would still exist for REDO_RENAME_TABLE. storage/maria/ma_loghandler.c: fail early if UNRECOVERABLE_ERROR storage/maria/ma_recovery.c: * new argument to maria_apply_log(): should it take checkpoints (at end of REDO phase and at the very end) or no. * moving the call to translog_next_LSN() into parse_checkpoint_record() ("hide the details"). * Refining an error detection for something which could happen if there is a checkpoint record in the log. * Using close_one_table() instead of maria_extra(HA_EXTRA_PREPARE_FOR_DROP|RENAME), as it looks safer, and also changing how close_one_table() works: it now limits itself to scanning all_tables[], thus having one loopp instead of two, which should be faster (as a result, it does not close tables not registered in this array, which is ok as there should not be any). storage/maria/ma_recovery.h: new parameter storage/maria/maria_read_log.c: update to new prototype
2007-10-08 19:08:25 +02:00
dfiles= kfiles= NULL;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pthread_mutex_destroy(&LOCK_checkpoint);
pthread_cond_destroy(&COND_checkpoint);
checkpoint_inited= FALSE;
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
DBUG_VOID_RETURN;
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/**
@brief dirty-page filtering criteria for MEDIUM checkpoint.
We flush data/index pages which have been dirty since the previous
checkpoint (this is the two-checkpoint rule: the REDO phase will not have
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
to start from earlier than the next-to-last checkpoint).
Bitmap pages are handled by _ma_bitmap_flush_all().
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
@param type Page's type
@param pageno Page's number
@param rec_lsn Page's rec_lsn
@param arg filter_param
*/
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
static enum pagecache_flush_filter_result
filter_flush_file_medium(enum pagecache_page_type type,
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
pgcache_page_no_t pageno __attribute__ ((unused)),
LSN rec_lsn, void *arg)
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
{
struct st_filter_param *param= (struct st_filter_param *)arg;
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
return (type == PAGECACHE_LSN_PAGE) &&
(cmp_translog_addr(rec_lsn, param->up_to_lsn) <= 0);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
}
/**
@brief dirty-page filtering criteria for FULL checkpoint.
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
We flush all dirty data/index pages.
Bitmap pages are handled by _ma_bitmap_flush_all().
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
@param type Page's type
@param pageno Page's number
@param rec_lsn Page's rec_lsn
@param arg filter_param
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
static enum pagecache_flush_filter_result
filter_flush_file_full(enum pagecache_page_type type,
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
pgcache_page_no_t pageno __attribute__ ((unused)),
LSN rec_lsn __attribute__ ((unused)),
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
void *arg __attribute__ ((unused)))
{
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
return (type == PAGECACHE_LSN_PAGE);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
}
/**
@brief dirty-page filtering criteria for background flushing thread.
We flush data/index pages which have been dirty since the previous
checkpoint (this is the two-checkpoint rule: the REDO phase will not have
to start from earlier than the next-to-last checkpoint), and no
bitmap pages. But we flush no more than a certain number of pages (to have
an even flushing, no write burst).
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
The reason to not flush bitmap pages is that they may not be in a flushable
state at this moment and we don't want to wait for them.
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
@param type Page's type
@param pageno Page's number
@param rec_lsn Page's rec_lsn
@param arg filter_param
*/
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
static enum pagecache_flush_filter_result
filter_flush_file_evenly(enum pagecache_page_type type,
pgcache_page_no_t pageno __attribute__ ((unused)),
LSN rec_lsn, void *arg)
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
{
struct st_filter_param *param= (struct st_filter_param *)arg;
if (unlikely(param->max_pages == 0)) /* all flushed already */
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
return FLUSH_FILTER_SKIP_ALL;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
if ((type == PAGECACHE_LSN_PAGE) &&
(cmp_translog_addr(rec_lsn, param->up_to_lsn) <= 0))
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
param->max_pages--;
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
return FLUSH_FILTER_OK;
}
WL#3071 Maria checkpoint Ability for flush_pagecache_blocks() to flush only certain pages of a file, as instructed by an option "filter" pointer-to-function argument; Checkpoint and background dirty page flushing use that to flush only pages which have been dirty for long enough and bitmap pages. Fix for a bug in flush_cached_blocks() (no idea if it could produce a bug in real life, but theoretically it is). Testing checkpoint in ma_test_recovery via ma_test1 and ma_test2. Background checkpoint & dirty pages flush thread is still disabled by default in ha_maria. mysql-test/r/maria.result: result update storage/maria/ha_maria.cc: blank after function comment storage/maria/ma_checkpoint.c: Using an enum instead of 0/1/2 (applying Sanja's review comments). The comment about "this is an horizon" can be removed as Sanja created translog_next_LSN() which parse_checkpoint_record() uses. Variables in ma_checkpoint_background() cannot be declared in the for() as their value must not be reset at each iteration! storage/maria/ma_pagecache.c: adding to flush_pagecache_blocks() optional arguments 'filter' (pointer to function) and 'filter_arg'; if filter!=NULL this function will be called for each block of the file and will reply if this block and following ones should be flushed or not (3 possible replies). Fixing a bug when flush_cached_blocks() skips a pinned page: it has to unset PCBLOCK_IN_FLUSH set by flush_pagecache_blocks_int(). storage/maria/ma_pagecache.h: flush_pagecache_blocks() is changed to take "filter" and "filter_arg" arguments. "filter", if it is not NULL, may return one value among enum pagecache_flush_filter_result. storage/maria/ma_recovery.c: open_count=0 when closing tables at the end of recovery. storage/maria/ma_test1.c: Optional checkpoints (-H#) at various stages (stages similar to --testflag), for testing of checkpoints. storage/maria/ma_test2.c: Optional checkpoints (-H#) at various stages (stages similar to -t), for testing of checkpoints. storage/maria/ma_test_recovery.expected: Result update: the results of the additional test run with -H# (checkpoints) are added here. They are exactly identical to without checkpoints except that the index's Root (printed by maria_chk) is more correct when using checkpoints. This is because checkpoint flushed the state, so it happens to be correct, while no-checkpoint does not flush the state, and recovery does not recover indexes so Root is never fixed. When we recover indices, this will go away. storage/maria/ma_test_recovery: We duplicate the loop of tests to add an additional run with checkpoints at various stages, to see if maria_read_log uses them fine.
2007-10-17 16:55:26 +02:00
return FLUSH_FILTER_SKIP_TRY_NEXT;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
}
/**
@brief Background thread which does checkpoints and flushes periodically.
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
Takes a checkpoint. After this, all pages dirty at the time of that
checkpoint are flushed evenly until it is time to take another checkpoint.
This ensures that the REDO phase starts at earliest (in LSN time) at the
next-to-last checkpoint record ("two-checkpoint rule").
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
@note MikaelR questioned why the same thread does two different jobs, the
risk could be that while a checkpoint happens no LRD flushing happens.
*/
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
pthread_handler_t ma_checkpoint_background(void *arg)
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
{
WL#3071 - Maria checkpoint * Preparation for having a background checkpoint thread: frequency of checkpoint taken by that thread is now configurable by the user: global variable maria_checkpoint_frequency, in seconds, default 30 (checkpoint every 30th second); 0 means no checkpoints (and thus no background thread, thus no background flushing, that will probably only be used for testing). * Don't take checkpoints in Recovery if it didn't do anything significant; thus no checkpoint after a clean shutdown/restart. The only checkpoint which is never skipped is the one at shutdown. * fix for a test failure (after-merge fix) include/maria.h: new variable mysql-test/suite/rpl/r/rpl_row_flsh_tbls.result: result update mysql-test/suite/rpl/t/rpl_row_flsh_tbls.test: position update (=after merge fix, as this position was already changed into 5.1 and not merged here, causing test to fail) storage/maria/ha_maria.cc: Checkpoint's frequency is now configurable by the user: global variable maria_checkpoint_frequency. Changing it on the fly requires us to shutdown/restart the background checkpoint thread, as the loop done in that thread assumes a constant checkpoint interval. Default value is 30: a checkpoint every 30 seconds (yes, I know, physicists will remind that it should be named "period" then). ha_maria now asks for a background checkpoint thread when it starts, but this is still overruled (disabled) in ma_checkpoint_init(). storage/maria/ma_checkpoint.c: Checkpoint's frequency is now configurable by the user: background thread takes a checkpoint every maria_checkpoint_interval-th second. If that variable is 0, no checkpoints are taken. Note, I will enable the background thread only in a later changeset. storage/maria/ma_recovery.c: Don't take checkpoints at the end of the REDO phase and at the end of Recovery if Recovery didn't make anything significant (didn't open any tables, didn't rollback any transactions). With this, after a clean shutdown, Recovery shouldn't take any checkpoint, which makes starting faster (we save a few fsync()s of the log and control file).
2007-10-09 10:38:31 +02:00
/** @brief At least this of log/page bytes written between checkpoints */
const uint checkpoint_min_activity= 2*1024*1024;
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
/*
If the interval could be changed by the user while we are in this thread,
it could be annoying: for example it could cause "case 2" to be executed
right after "case 0", thus having 'dfile' unset. So the thread cares only
about the interval's value when it started.
*/
const ulong interval= (ulong)arg;
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
uint sleeps, sleep_time;
TRANSLOG_ADDRESS log_horizon_at_last_checkpoint=
translog_get_horizon();
ulonglong pagecache_flushes_at_last_checkpoint=
maria_pagecache->global_cache_write;
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
uint pages_bunch_size;
struct st_filter_param filter_param;
PAGECACHE_FILE *dfile; /**< data file currently being flushed */
PAGECACHE_FILE *kfile; /**< index file currently being flushed */
LINT_INIT(kfile);
LINT_INIT(dfile);
LINT_INIT(pages_bunch_size);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
my_thread_init();
DBUG_PRINT("info",("Maria background checkpoint thread starts"));
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
DBUG_ASSERT(interval > 0);
/*
Recovery ended with all tables closed and a checkpoint: no need to take
one immediately.
*/
sleeps= 1;
pages_to_flush_before_next_checkpoint= 0;
for(;;) /* iterations of checkpoints and dirty page flushing */
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
{
#if 0 /* good for testing, to do a lot of checkpoints, finds a lot of bugs */
sleeps=0;
#endif
struct timespec abstime;
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
switch (sleeps % interval)
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
{
case 0:
/*
With background flushing evenly distributed over the time
between two checkpoints, we should have only little flushing to do
in the checkpoint.
*/
/*
No checkpoint if little work of interest for recovery was done
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
since last checkpoint. Such work includes log writing (lengthens
recovery, checkpoint would shorten it), page flushing (checkpoint
would decrease the amount of read pages in recovery).
WL#3071 - Maria checkpoint * Preparation for having a background checkpoint thread: frequency of checkpoint taken by that thread is now configurable by the user: global variable maria_checkpoint_frequency, in seconds, default 30 (checkpoint every 30th second); 0 means no checkpoints (and thus no background thread, thus no background flushing, that will probably only be used for testing). * Don't take checkpoints in Recovery if it didn't do anything significant; thus no checkpoint after a clean shutdown/restart. The only checkpoint which is never skipped is the one at shutdown. * fix for a test failure (after-merge fix) include/maria.h: new variable mysql-test/suite/rpl/r/rpl_row_flsh_tbls.result: result update mysql-test/suite/rpl/t/rpl_row_flsh_tbls.test: position update (=after merge fix, as this position was already changed into 5.1 and not merged here, causing test to fail) storage/maria/ha_maria.cc: Checkpoint's frequency is now configurable by the user: global variable maria_checkpoint_frequency. Changing it on the fly requires us to shutdown/restart the background checkpoint thread, as the loop done in that thread assumes a constant checkpoint interval. Default value is 30: a checkpoint every 30 seconds (yes, I know, physicists will remind that it should be named "period" then). ha_maria now asks for a background checkpoint thread when it starts, but this is still overruled (disabled) in ma_checkpoint_init(). storage/maria/ma_checkpoint.c: Checkpoint's frequency is now configurable by the user: background thread takes a checkpoint every maria_checkpoint_interval-th second. If that variable is 0, no checkpoints are taken. Note, I will enable the background thread only in a later changeset. storage/maria/ma_recovery.c: Don't take checkpoints at the end of the REDO phase and at the end of Recovery if Recovery didn't make anything significant (didn't open any tables, didn't rollback any transactions). With this, after a clean shutdown, Recovery shouldn't take any checkpoint, which makes starting faster (we save a few fsync()s of the log and control file).
2007-10-09 10:38:31 +02:00
In case of one short statement per minute (very low load), we don't
want to checkpoint every minute, hence the positive
checkpoint_min_activity.
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
*/
if (((translog_get_horizon() - log_horizon_at_last_checkpoint) +
(maria_pagecache->global_cache_write -
pagecache_flushes_at_last_checkpoint) *
maria_pagecache->block_size) < checkpoint_min_activity)
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
{
/* don't take checkpoint, so don't know what to flush */
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pages_to_flush_before_next_checkpoint= 0;
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
sleep_time= interval;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
break;
}
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
sleep_time= 1;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
ma_checkpoint_execute(CHECKPOINT_MEDIUM, TRUE);
/*
Snapshot this kind of "state" of the engine. Note that the value below
is possibly greater than last_checkpoint_lsn.
*/
log_horizon_at_last_checkpoint= translog_get_horizon();
pagecache_flushes_at_last_checkpoint=
maria_pagecache->global_cache_write;
/*
If the checkpoint above succeeded it has set d|kfiles and
d|kfiles_end. If is has failed, it has set
pages_to_flush_before_next_checkpoint to 0 so we will skip flushing
and sleep until the next checkpoint.
*/
break;
case 1:
/* set up parameters for background page flushing */
filter_param.up_to_lsn= last_checkpoint_lsn;
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
pages_bunch_size= pages_to_flush_before_next_checkpoint / interval;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
dfile= dfiles;
kfile= kfiles;
/* fall through */
default:
if (pages_bunch_size > 0)
{
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
DBUG_PRINT("checkpoint",
("Maria background checkpoint thread: %u pages",
pages_bunch_size));
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* flush a bunch of dirty pages */
filter_param.max_pages= pages_bunch_size;
while (dfile != dfiles_end)
{
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
/*
We use FLUSH_KEEP_LAZY: if a file is already in flush, it's
smarter to move to the next file than wait for this one to be
completely flushed, which may take long.
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
int res=
flush_pagecache_blocks_with_filter(maria_pagecache,
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
dfile, FLUSH_KEEP_LAZY,
filter_flush_file_evenly,
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
&filter_param);
if (unlikely(res & PCFLUSH_ERROR))
WL#3071 Maria checkpoint, WL#3072 Maria recovery instead of fprintf(stderr) when a task (with no user connected) gets an error, use my_printf_error(). Flags ME_JUST_WARNING and ME_JUST_INFO added to my_error()/my_printf_error(), which pass it to my_message_sql() which is modified to call the appropriate sql_print_*(). This way recovery can signal its start and end with [Note] and not [ERROR] (but failure with [ERROR]). Recovery's detailed progress (percents etc) still uses stderr as they have to stay on one single line. sql_print_error() changed to use my_progname_short (nicer display). mysql-test-run.pl --gdb/--ddd does not run mysqld, because a breakpoint in mysql_parse is too late to debug startup problems; instead, dev should set the breakpoints it wants and then "run" ("r"). include/my_sys.h: new flags to tell error_handler_hook that this is not an error but an information or warning mysql-test/mysql-test-run.pl: when running with --gdb/--ddd to debug mysqld, breaking at mysql_parse is too late to debug startup problems; now, it does not run mysqld, does not set breakpoints, developer can set as early breakpoints as it wants and is responsible for typing "run" (or "r") mysys/my_init.c: set my_progname_short mysys/my_static.c: my_progname_short added sql/mysqld.cc: * my_message_sql() can now receive info or warning, not only error; this allows mysys to tell the user (or the error log if no user) about an info or warning. Used from Maria. * plugins (or engines like Maria) may want to call my_error(), so set up the error handler hook (my_message_sql) before initializing plugins; otherwise they get my_message_no_curses which is less integrated into mysqld (is just fputs()) * using my_progname_short instead of my_progname, in my_message_sql() (less space on screen) storage/maria/ma_checkpoint.c: fprintf(stderr) -> ma_message_no_user() storage/maria/ma_checkpoint.h: function for any Maria task, not connected to a user (example: checkpoint, recovery; soon could be deleted records purger) to report a message (calls my_printf_error() which, when inside ha_maria, leads to sql_print_*(), and when outside, leads to my_message_no_curses i.e. stderr). storage/maria/ma_recovery.c: To tell that recovery starts and ends we use ma_message_no_user() (sql_print_*() in practice). Detailed progress info still uses stderr as sql_print() cannot put several messages on one line. 071116 18:42:16 [Note] mysqld: Maria engine: starting recovery recovered pages: 0% 67% 100% (0.0 seconds); transactions to roll back: 1 0 (0.0 seconds); tables to flush: 1 0 (0.0 seconds); 071116 18:42:16 [Note] mysqld: Maria engine: recovery done storage/maria/maria_chk.c: my_progname_short moved to mysys storage/maria/maria_read_log.c: my_progname_short moved to mysys storage/myisam/myisamchk.c: my_progname_short moved to mysys
2007-11-16 17:09:51 +01:00
ma_message_no_user(0, "background data page flush failed");
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
if (filter_param.max_pages == 0) /* bunch all flushed, sleep */
break; /* and we will continue with the same file */
dfile++; /* otherwise all this file is flushed, move to next file */
2007-11-11 19:01:53 +01:00
/*
MikaelR noted that he observed that Linux's file cache may never
fsync to disk until this cache is full, at which point it decides
to empty the cache, making the machine very slow. A solution was
to fsync after writing 2 MB. So we might want to fsync() here if
we wrote enough pages.
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
}
while (kfile != kfiles_end)
{
int res=
flush_pagecache_blocks_with_filter(maria_pagecache,
kfile, FLUSH_KEEP_LAZY,
filter_flush_file_evenly,
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
&filter_param);
if (unlikely(res & PCFLUSH_ERROR))
WL#3071 Maria checkpoint, WL#3072 Maria recovery instead of fprintf(stderr) when a task (with no user connected) gets an error, use my_printf_error(). Flags ME_JUST_WARNING and ME_JUST_INFO added to my_error()/my_printf_error(), which pass it to my_message_sql() which is modified to call the appropriate sql_print_*(). This way recovery can signal its start and end with [Note] and not [ERROR] (but failure with [ERROR]). Recovery's detailed progress (percents etc) still uses stderr as they have to stay on one single line. sql_print_error() changed to use my_progname_short (nicer display). mysql-test-run.pl --gdb/--ddd does not run mysqld, because a breakpoint in mysql_parse is too late to debug startup problems; instead, dev should set the breakpoints it wants and then "run" ("r"). include/my_sys.h: new flags to tell error_handler_hook that this is not an error but an information or warning mysql-test/mysql-test-run.pl: when running with --gdb/--ddd to debug mysqld, breaking at mysql_parse is too late to debug startup problems; now, it does not run mysqld, does not set breakpoints, developer can set as early breakpoints as it wants and is responsible for typing "run" (or "r") mysys/my_init.c: set my_progname_short mysys/my_static.c: my_progname_short added sql/mysqld.cc: * my_message_sql() can now receive info or warning, not only error; this allows mysys to tell the user (or the error log if no user) about an info or warning. Used from Maria. * plugins (or engines like Maria) may want to call my_error(), so set up the error handler hook (my_message_sql) before initializing plugins; otherwise they get my_message_no_curses which is less integrated into mysqld (is just fputs()) * using my_progname_short instead of my_progname, in my_message_sql() (less space on screen) storage/maria/ma_checkpoint.c: fprintf(stderr) -> ma_message_no_user() storage/maria/ma_checkpoint.h: function for any Maria task, not connected to a user (example: checkpoint, recovery; soon could be deleted records purger) to report a message (calls my_printf_error() which, when inside ha_maria, leads to sql_print_*(), and when outside, leads to my_message_no_curses i.e. stderr). storage/maria/ma_recovery.c: To tell that recovery starts and ends we use ma_message_no_user() (sql_print_*() in practice). Detailed progress info still uses stderr as sql_print() cannot put several messages on one line. 071116 18:42:16 [Note] mysqld: Maria engine: starting recovery recovered pages: 0% 67% 100% (0.0 seconds); transactions to roll back: 1 0 (0.0 seconds); tables to flush: 1 0 (0.0 seconds); 071116 18:42:16 [Note] mysqld: Maria engine: recovery done storage/maria/maria_chk.c: my_progname_short moved to mysys storage/maria/maria_read_log.c: my_progname_short moved to mysys storage/myisam/myisamchk.c: my_progname_short moved to mysys
2007-11-16 17:09:51 +01:00
ma_message_no_user(0, "background index page flush failed");
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
if (filter_param.max_pages == 0) /* bunch all flushed, sleep */
break; /* and we will continue with the same file */
kfile++; /* otherwise all this file is flushed, move to next file */
}
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
sleep_time= 1;
}
else
{
/* Can directly sleep until the next checkpoint moment */
sleep_time= interval - (sleeps % interval);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
}
}
pthread_mutex_lock(&LOCK_checkpoint);
if (checkpoint_thread_die == 1)
break;
#if 0 /* good for testing, to do a lot of checkpoints, finds a lot of bugs */
pthread_mutex_unlock(&LOCK_checkpoint);
WL#3072 - Maria recovery. * Recovery of the table's live checksum (CREATE TABLE ... CHECKSUM=1) is achieved in this patch. The table's live checksum (info->s->state.state.checksum) is updated in inwrite_rec_hook's under the log mutex when writing UNDO_ROW_INSERT|UPDATE|DELETE and REDO_DELETE_ALL. The checksum variation caused by the operation is stored in these UNDOs, so that the REDO phase, when it sees such UNDOs, can update the live checksum if it is older (state.is_of_lsn is lower) than the record. It is also used, as a nice add-on with no cost, to do less row checksum computation during the UNDO phase (as we have it in the record already). Doing this work, it became pressing to move in-write hooks (write_hook_for_redo() et al) to ma_blockrec.c. The 'parts' argument of inwrite_rec_hook is unpredictable (it comes mangled at this stage, for example by LSN compression) so it is replaced by a 'void* hook_arg', which is used to pass down information, currently only to write_hook_for_clr_end() (previous undo_lsn and type of undone record). * If from ha_maria, we print to stderr how many seconds (with one fractional digit) the REDO phase took, same for UNDO phase and for final table close. Just to give an indication for debugging and maybe also for Support. storage/maria/ha_maria.cc: question for Monty storage/maria/ma_blockrec.c: * log in-write hooks (write_hook_for_redo() etc) move from ma_loghandler.c to here; this is natural: the hooks are coupled to their callers (functions in ma_blockrec.c). * translog_write_record() now has a new argument "hook_arg"; using it to pass down to write_hook_for_clr_end() the transaction's previous_undo_lsn and the type of the being undone record, and also to pass down to all UNDOs the live checksum variation caused by the operation. * If table has live checksum, store in UNDO_ROW_INSERT|UPDATE|DELETE and in CLR_END the checksum variation ("delta") caused by the operation. For example if a DELETE caused the table's live checksum to change from 123 to 456, we store in the UNDO_ROW_DELETE, in 4 bytes, the value 333 (456-123). * Instead of hard-coded "1" as length of the place where we store the undone record's type in CLR_END, use a symbol CLR_TYPE_STORE_SIZE; use macros clr_type_store and clr_type_korr. * write_block_record() has a new parameter 'old_record_checksum' which is the pre-computed checksum of old_record; that value is used to update the table's live checksum when writing UNDO_ROW_UPDATE|CLR_END. * In allocate_write_block_record(), if we are executing UNDO_ROW_DELETE the row's checksum is already computed. * _ma_update_block_record2() now expect the new row's checksum into cur_row.checksum (was already true) and the old row's checksum into new_row.checksum (that's new). Its two callers, maria_update() and _ma_apply_undo_row_update(), honour this. * When executing an UNDO_ROW_INSERT|UPDATE|DELETE in UNDO phase, pick up the checksum delta from the log record. It is then used to update the table's live checksum when writing CLR_END, and saves us a computation of record. storage/maria/ma_blockrec.h: in-write hooks move from ma_loghandler.c storage/maria/ma_check.c: more straightforward size of buffer storage/maria/ma_checkpoint.c: <= is enough storage/maria/ma_commit.c: new prototype of translog_write_record() storage/maria/ma_create.c: new prototype of translog_write_record() storage/maria/ma_delete.c: The row's checksum must be computed before calling(*delete_record)(), not after, because it must be known inside _ma_delete_block_record() (to update the table's live checksum when writing UNDO_ROW_DELETE). If deleting from a transactional table, live checksum was already updated when writing UNDO_ROW_DELETE. storage/maria/ma_delete_all.c: @todo is now done (in ma_loghandler.c) storage/maria/ma_delete_table.c: new prototype of translog_write_record() storage/maria/ma_loghandler.c: * in-write hooks move to ma_blockrec.c. * translog_write_record() gets a new argument 'hook_arg' which is passed down to pre|inwrite_rec_hook. It is more useful that 'parts' for those hooks, because when those hooks are called, 'parts' has possibly been mangled (like with LSN compression) and is so unpredictable. * fix for compiler warning (unused buffer_start when compiling without debug support) * Because checksum delta is stored into UNDO_ROW_INSERT|UPDATE|DELETE and CLR_END, but only if the table has live checksum, these records are not PSEUDOFIXEDLENGTH anymore, they are now VARIABLE_LENGTH (their length is X if no live checksum and X+4 otherwise). * add an inwrite_rec_hook for UNDO_ROW_UPDATE, which updates the table's live checksum. Update it also in hooks of UNDO_ROW_INSERT| DELETE and REDO_DELETE_ALL and CLR_END. * Bugfix: when reading a record in translog_read_record(), it happened that "length" became negative, because the function assumed that the record extended beyond the page's end, whereas it may be shorter. storage/maria/ma_loghandler.h: * Instead of hard-coded "1" and "4", use symbols and macros to store/retrieve the type of record which the CLR_END corresponds to, and the checksum variation caused by the operation which logs the record * translog_write_record() gets a new argument 'hook_arg' which is passed down to pre|inwrite_rec_hook. It is more useful that 'parts' for those hooks, because when those hooks are called, 'parts' has possibly been mangled (like with LSN compression) and is so unpredictable. storage/maria/ma_open.c: fix for "empty body in if() statement" (when compiling without safemutex) storage/maria/ma_pagecache.c: <= is enough storage/maria/ma_recovery.c: * print the time that each recovery phase (REDO/UNDO/flush) took; this is enabled only when recovering from ha_maria. Is it printed n seconds with a fractional part of one digit (like 123.4 seconds). * In the REDO phase, update the table's live checksum by using the checksum delta stored in UNDO_ROW_INSERT|DELETE|UPDATE and CLR_END. Update it too when seeing REDO_DELETE_ALL. * In the UNDO phase, when executing UNDO_ROW_INSERT, if the table does not have live checksum then reading the record's header (as done by the master loop of run_undo_phase()) is enough; otherwise we do a translog_read_record() to have the checksum delta ready for _ma_apply_undo_row_insert(). * When at the end of the REDO phase we notice that there is an unfinished group of REDOs, don't assert in debug binaries, as I verified that it can happen in real life (with kill -9) * removing ' in #error as it confuses gcc3 storage/maria/ma_rename.c: new prototype of translog_write_record() storage/maria/ma_test_recovery.expected: Change in output of ma_test_recovery: now all live checksums of original tables equal those of tables recreated by the REDO phase and those of tables fixed by the UNDO phase. I.e. recovery of the live checksum looks like working (which was after all the only goal of this changeset). I checked by hand that it's not just all live checksums which are now 0 and that's why they match. They are the old values like 3757530372. maria.test has hard-coded checksum values in its result file so checks this too. storage/maria/ma_update.c: * It's useless to put up HA_STATE_CHANGED in 'key_changed', as we put up HA_STATE_CHANGED in info->update anyway. * We need to compute the old and new rows' checksum before calling (*update_record)(), as checksum delta must be known when logging UNDO_ROW_UPDATE which is done by _ma_update_block_record(). Note that some functions change the 'newrec' record (at least _ma_check_unique() does) so we cannot move the checksum computation too early in the function. storage/maria/ma_write.c: If inserting into a transactional table, live's checksum was already updated when writing UNDO_ROW_INSERT. The multiplication is a trick to save an if(). storage/maria/unittest/ma_test_loghandler-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_first_lsn-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_max_lsn-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_multigroup-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_multithread-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_noflush-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_pagecache-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_purge-t.c: new prototype of translog_write_record() storage/myisam/sort.c: fix for compiler warnings in pushbuild (write_merge_key* functions didn't have their declaration match MARIA_HA::write_key).
2007-10-02 18:02:09 +02:00
my_sleep(100000); /* a tenth of a second */
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pthread_mutex_lock(&LOCK_checkpoint);
#else
/* To have a killable sleep, we use timedwait like our SQL GET_LOCK() */
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
DBUG_PRINT("info", ("sleeping %u seconds", sleep_time));
set_timespec(abstime, sleep_time);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pthread_cond_timedwait(&COND_checkpoint, &LOCK_checkpoint, &abstime);
#endif
if (checkpoint_thread_die == 1)
break;
pthread_mutex_unlock(&LOCK_checkpoint);
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
sleeps+= sleep_time;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
}
pthread_mutex_unlock(&LOCK_checkpoint);
DBUG_PRINT("info",("Maria background checkpoint thread ends"));
WL#3072 - Maria Recovery * to honour WAL we now force the whole log when flushing a bitmap page. * ability to intentionally crash in various places for recovery testing * bugfix (dirty pages list found in checkpoint record was ignored) * smaller checkpoint record * misc small cleanups and comments mysql-test/include/maria_empty_logs.inc: maria-purge.test creates ~11 logs, remove them all mysql-test/r/maria-recovery-bitmap.result: result is good; without the _ma_bitmap_get_log_address() call, we got check error Bitmap at 0 has pages reserved outside of data file length mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery-bitmap.test: enable test of "bitmap-flush should flush whole log otherwise corrupted data file (bitmap ahead of data pages)". mysql-test/t/maria-recovery.test: test of checkpoint sql/sql_table.cc: comment storage/maria/ha_maria.cc: _ma_reenable_logging_for_table() now includes file->trn=0. At the end of repair() we don't need to re-enable logging, it is done already by caller (like copy_data_between_tables()); it sounds strange that this function could decide to re-enable, it should be up to caller who knows what other operations it plans. Removing this line led to assertion failure in maria_lock_database(F_UNLCK), fixed by removing the assertion: maria_lock_database() is here called in a context where F_UNLCK does not make the table visible to others so assertion is excessive, and external_lock() is already designed to honour the asserted condition. Ability to crash at the end of bulk insert when indices have been enabled. storage/maria/ma_bitmap.c: Better use pagecache_file_init() than set pagecache callbacks directly; and a new function to set those callbacks for bitmap so that we can reuse it. _ma_bitmap_get_log_address() is a pagecache get_log_address callback which causes the whole log to be flushed when a bitmap page is flushed by the page cache. This was required by WAL. storage/maria/ma_blockrec.c: get_log_address pagecache callback for data (non bitmap) pages: just reads the LSN from the page's content, like was hard-coded before in ma_pagecache.c. storage/maria/ma_blockrec.h: functions which need to be exported storage/maria/ma_check.c: create_new_data_handle() can be static. Ability to crash after rebuilding the index in OPTIMIZE, in REPAIR. my_lock() implemented already. storage/maria/ma_checkpoint.c: As MARIA_SHARE* is now accessible to pagecache_collect_changed_blocks_LSN(), we don't need to store kfile/dfile descriptors in checkpoint record, 2-byte-id of the table plus one byte to say if this is data or index file is enough. So we go from 4+4 bytes per table down to 2+1. storage/maria/ma_commit.c: removing duplicate functions (see _ma_tmp_disable_logging_for_table()) storage/maria/ma_extra.c: Monty fixed storage/maria/ma_key_recover.c: comment storage/maria/ma_locking.c: Sometimes other code does funny things with maria_lock_database(), like ha_maria::repair() calling it at start and end without going through ha_maria::external_lock(). So it happens that maria_lock_database() is called with now_transactional!=born_transactional. storage/maria/ma_loghandler.c: update to new prototype storage/maria/ma_open.c: set_data|index_pagecache_callbacks() need to be exported as they are now called when disabling/enabling transactionality. storage/maria/ma_pagecache.c: Removing PAGE_LSN_OFFSET, as much of the code relies on it being 0 anyway (let's not give impression we can just change this constant). When flushing a page to disk, call the get_log_address callback to know up to which LSN the log should be flushed. As we now can access MARIA_SHARE* we can know share->id and store it into the checkpoint record; we thus go from 4 bytes per dirty page to 2+1. storage/maria/ma_pagecache.h: get_log_address callback storage/maria/ma_panic.c: No reason to reset pagecache callbacks in HA_PANIC_READ: all we do is reopen files if they were closed; callbacks should be in place already as 'info' exists; we just want to modify the file descriptors, not the full PAGECACHE_FILE structure. If we open data file and it was closed, share->bitmap.file needs to be set. Note that the modified code is disabled anyway. storage/maria/ma_recovery.c: Checkpoint record does not contain kfile/dfile descriptors anymore so code can be simplified. Hash key in all_dirty_pages is not made from file_descriptor & pageno anymore, but index_or_data & table-short-id & pageno. If a table's create_rename_lsn is higher than record's LSN, we skip the table and don't fail if it's corrupted (because the LSNs say that we don't have to look at this table). If a table is skipped (for example due to create_rename_lsn), its UNDOs still cause undo_lsn to advance; this is so that if later we notice the transaction has to rollback we fail (as table should not be skipped in this case). Fixing a bug: the dirty_pages list was never used, because the LSN below which it was used was the minimum rec_lsn of dirty pages! It is now the min(checkpoint_start_log_horizon, min(trn's rec_lsn)). When we disable/reenable transactionality, we modify pagecache callbacks (needed for example for get_log_address: changing share->page_type is not enough anymore). storage/maria/ma_write.c: 'records' and 'checksum' are protected: they are updated under log's mutex in write-hooks when UNDO is written. storage/maria/maria_chk.c: remove use of duplicate functions. storage/maria/maria_def.h: set_data|index_pagecache_callbacks() need to be exported; _ma_reenable_logging_for_table() changes to a real function. storage/maria/unittest/ma_pagecache_consist.c: new prototype storage/maria/unittest/ma_pagecache_single.c: new prototype storage/maria/unittest/ma_test_loghandler_pagecache-t.c: new prototype
2007-12-30 21:32:07 +01:00
{
CHECKPOINT_LEVEL level= CHECKPOINT_FULL;
/*
That's the final one, which guarantees that a clean shutdown always ends
with a checkpoint.
*/
DBUG_EXECUTE_IF("maria_checkpoint_indirect", level= CHECKPOINT_INDIRECT;);
ma_checkpoint_execute(level, FALSE);
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pthread_mutex_lock(&LOCK_checkpoint);
checkpoint_thread_die= 2; /* indicate that we are dead */
/* wake up ma_checkpoint_end() which may be waiting for our death */
pthread_cond_broadcast(&COND_checkpoint);
/* broadcast was inside unlock because ma_checkpoint_end() destroys mutex */
pthread_mutex_unlock(&LOCK_checkpoint);
my_thread_end();
return 0;
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/**
@brief Allocates buffer and stores in it some info about open tables,
does some flushing on those.
Does the allocation because the caller cannot know the size itself.
Memory freeing is to be done by the caller (if the "str" member of the
LEX_STRING is not NULL).
The caller is taking a checkpoint.
@param[out] str pointer to where the allocated buffer,
and its size, will be put; buffer will be filled
with info about open tables
@param checkpoint_start_log_horizon Of the in-progress checkpoint
record.
@return Operation status
@retval 0 OK
@retval 1 Error
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
static int collect_tables(LEX_STRING *str, LSN checkpoint_start_log_horizon)
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
MARIA_SHARE **distinct_shares= NULL;
char *ptr;
uint error= 1, sync_error= 0, nb, nb_stored, i;
my_bool unmark_tables= TRUE;
uint total_names_length;
LIST *pos; /**< to iterate over open tables */
struct st_state_copy {
uint index;
MARIA_STATE_INFO state;
};
struct st_state_copy *state_copies= NULL, /**< fixed-size cache of states */
*state_copies_end, /**< cache ends here */
*state_copy; /**< iterator in cache */
TRANSLOG_ADDRESS state_copies_horizon; /**< horizon of states' _copies_ */
struct st_filter_param filter_param;
PAGECACHE_FLUSH_FILTER filter;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
DBUG_ENTER("collect_tables");
LINT_INIT(state_copies_horizon);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* let's make a list of distinct shares */
pthread_mutex_lock(&THR_LOCK_maria);
for (nb= 0, pos= maria_open_list; pos; pos= pos->next)
{
MARIA_HA *info= (MARIA_HA*)pos->data;
MARIA_SHARE *share= info->s;
/* the first three variables below can never change */
if (share->base.born_transactional && !share->temporary &&
share->mode != O_RDONLY &&
!(share->in_checkpoint & MARIA_CHECKPOINT_SEEN_IN_LOOP))
{
/*
Why we didn't take intern_lock above: table had in_checkpoint==0 so no
thread could set in_checkpoint. And no thread needs to know that we
are setting in_checkpoint, because only maria_close() needs it and
cannot run now as we hold THR_LOCK_maria.
*/
/*
This table is relevant for checkpoint and not already seen. Mark it,
so that it is not seen again in the loop.
*/
nb++;
DBUG_ASSERT(share->in_checkpoint == 0);
/* This flag ensures that we count only _distinct_ shares. */
share->in_checkpoint= MARIA_CHECKPOINT_SEEN_IN_LOOP;
}
}
if (unlikely((distinct_shares=
(MARIA_SHARE **)my_malloc(nb * sizeof(MARIA_SHARE *),
MYF(MY_WME))) == NULL))
goto err;
for (total_names_length= 0, i= 0, pos= maria_open_list; pos; pos= pos->next)
{
MARIA_HA *info= (MARIA_HA*)pos->data;
MARIA_SHARE *share= info->s;
if (share->in_checkpoint & MARIA_CHECKPOINT_SEEN_IN_LOOP)
{
distinct_shares[i++]= share;
/*
With this we prevent the share from going away while we later flush
and force it without holding THR_LOCK_maria. For example if the share
could be my_free()d by maria_close() we would have a problem when we
access it to flush the table. We "pin" the share pointer.
And we also take down MARIA_CHECKPOINT_SEEN_IN_LOOP, so that it is
not seen again in the loop.
*/
share->in_checkpoint= MARIA_CHECKPOINT_LOOKS_AT_ME;
/** @todo avoid strlen() */
total_names_length+= strlen(share->open_file_name);
}
}
DBUG_ASSERT(i == nb);
pthread_mutex_unlock(&THR_LOCK_maria);
DBUG_PRINT("info",("found %u table shares", nb));
str->length=
4 + /* number of tables */
(2 + /* short id */
LSN_STORE_SIZE + /* first_log_write_at_lsn */
1 /* end-of-name 0 */
) * nb + total_names_length;
if (unlikely((str->str= my_malloc(str->length, MYF(MY_WME))) == NULL))
goto err;
ptr= str->str;
ptr+= 4; /* real number of stored tables is not yet know */
/* only possible checkpointer, so can do the read below without mutex */
filter_param.up_to_lsn= last_checkpoint_lsn;
switch(checkpoint_in_progress)
{
case CHECKPOINT_MEDIUM:
filter= &filter_flush_file_medium;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
break;
case CHECKPOINT_FULL:
filter= &filter_flush_file_full;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
break;
case CHECKPOINT_INDIRECT:
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
filter= NULL;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
break;
default:
DBUG_ASSERT(0);
goto err;
}
/*
The principle of reading/writing the state below is explained in
ma_recovery.c, look for "Recovery of the state".
*/
#define STATE_COPIES 1024
state_copies= (struct st_state_copy *)
my_malloc(STATE_COPIES * sizeof(struct st_state_copy), MYF(MY_WME));
dfiles= (PAGECACHE_FILE *)my_realloc((uchar *)dfiles,
/* avoid size of 0 for my_realloc */
max(1, nb) * sizeof(PAGECACHE_FILE),
MYF(MY_WME | MY_ALLOW_ZERO_PTR));
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
kfiles= (PAGECACHE_FILE *)my_realloc((uchar *)kfiles,
/* avoid size of 0 for my_realloc */
max(1, nb) * sizeof(PAGECACHE_FILE),
MYF(MY_WME | MY_ALLOW_ZERO_PTR));
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
if (unlikely((state_copies == NULL) ||
(dfiles == NULL) || (kfiles == NULL)))
goto err;
state_copy= state_copies_end= NULL;
dfiles_end= dfiles;
kfiles_end= kfiles;
for (nb_stored= 0, i= 0; i < nb; i++)
{
MARIA_SHARE *share= distinct_shares[i];
PAGECACHE_FILE kfile, dfile;
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
my_bool ignore_share;
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
if (!(share->in_checkpoint & MARIA_CHECKPOINT_LOOKS_AT_ME))
{
/* No need for a mutex to read the above, only us can write this flag */
continue;
}
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
/**
@todo We should not look at tables which didn't change since last
checkpoint.
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
DBUG_PRINT("info",("looking at table '%s'", share->open_file_name));
if (state_copy == state_copies_end) /* we have no more cached states */
{
/*
Collect and cache a bunch of states. We do this for many states at a
time, to not lock/unlock the log's lock too often.
*/
uint j, bound= min(nb, i + STATE_COPIES);
state_copy= state_copies;
/* part of the state is protected by log's lock */
translog_lock();
state_copies_horizon= translog_get_horizon_no_lock();
for (j= i; j < bound; j++)
{
MARIA_SHARE *share2= distinct_shares[j];
if (!(share2->in_checkpoint & MARIA_CHECKPOINT_LOOKS_AT_ME))
continue;
state_copy->index= j;
state_copy->state= share2->state; /* we copy the state */
state_copy++;
/*
data_file_length is not updated under log's lock by the bitmap
code, but writing a wrong data_file_length is ok: a next
maria_close() will correct it; if we crash before, Recovery will
set it to the true physical size.
*/
}
translog_unlock();
/**
We are going to flush these states.
Before, all records describing how to undo such state must be
in the log (WAL). Usually this means UNDOs. In the special case of
data|key_file_length, recovery just needs to open the table to fix the
length, so any LOGREC_FILE_ID/REDO/UNDO allowing recovery to
understand it must open a table, is enough; so as long as
data|key_file_length is updated after writing any log record it's ok:
if we copied new value above, it means the record was before
state_copies_horizon and we flush such record below.
Apart from data|key_file_length which are easily recoverable from the
real file's size, all other state members must be updated only when
writing the UNDO; otherwise, if updated before, if their new value is
flushed by a checkpoint and there is a crash before UNDO is written,
their REDO group will be missing or at least incomplete and skipped
by recovery, so bad state value will stay. For example, setting
key_root before writing the UNDO: the table would have old index
pages (they were pinned at time of crash) and a new, thus wrong,
key_root.
@todo RECOVERY BUG check that all code honours that.
*/
if (translog_flush(state_copies_horizon))
goto err;
/* now we have cached states and they are WAL-safe*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
state_copies_end= state_copy;
state_copy= state_copies;
}
/* locate our state among these cached ones */
for ( ; state_copy->index != i; state_copy++)
DBUG_ASSERT(state_copy < state_copies_end);
/* OS file descriptors are ints which we stored in 4 bytes */
WL#3072 - Maria recovery. * Recovery of the table's live checksum (CREATE TABLE ... CHECKSUM=1) is achieved in this patch. The table's live checksum (info->s->state.state.checksum) is updated in inwrite_rec_hook's under the log mutex when writing UNDO_ROW_INSERT|UPDATE|DELETE and REDO_DELETE_ALL. The checksum variation caused by the operation is stored in these UNDOs, so that the REDO phase, when it sees such UNDOs, can update the live checksum if it is older (state.is_of_lsn is lower) than the record. It is also used, as a nice add-on with no cost, to do less row checksum computation during the UNDO phase (as we have it in the record already). Doing this work, it became pressing to move in-write hooks (write_hook_for_redo() et al) to ma_blockrec.c. The 'parts' argument of inwrite_rec_hook is unpredictable (it comes mangled at this stage, for example by LSN compression) so it is replaced by a 'void* hook_arg', which is used to pass down information, currently only to write_hook_for_clr_end() (previous undo_lsn and type of undone record). * If from ha_maria, we print to stderr how many seconds (with one fractional digit) the REDO phase took, same for UNDO phase and for final table close. Just to give an indication for debugging and maybe also for Support. storage/maria/ha_maria.cc: question for Monty storage/maria/ma_blockrec.c: * log in-write hooks (write_hook_for_redo() etc) move from ma_loghandler.c to here; this is natural: the hooks are coupled to their callers (functions in ma_blockrec.c). * translog_write_record() now has a new argument "hook_arg"; using it to pass down to write_hook_for_clr_end() the transaction's previous_undo_lsn and the type of the being undone record, and also to pass down to all UNDOs the live checksum variation caused by the operation. * If table has live checksum, store in UNDO_ROW_INSERT|UPDATE|DELETE and in CLR_END the checksum variation ("delta") caused by the operation. For example if a DELETE caused the table's live checksum to change from 123 to 456, we store in the UNDO_ROW_DELETE, in 4 bytes, the value 333 (456-123). * Instead of hard-coded "1" as length of the place where we store the undone record's type in CLR_END, use a symbol CLR_TYPE_STORE_SIZE; use macros clr_type_store and clr_type_korr. * write_block_record() has a new parameter 'old_record_checksum' which is the pre-computed checksum of old_record; that value is used to update the table's live checksum when writing UNDO_ROW_UPDATE|CLR_END. * In allocate_write_block_record(), if we are executing UNDO_ROW_DELETE the row's checksum is already computed. * _ma_update_block_record2() now expect the new row's checksum into cur_row.checksum (was already true) and the old row's checksum into new_row.checksum (that's new). Its two callers, maria_update() and _ma_apply_undo_row_update(), honour this. * When executing an UNDO_ROW_INSERT|UPDATE|DELETE in UNDO phase, pick up the checksum delta from the log record. It is then used to update the table's live checksum when writing CLR_END, and saves us a computation of record. storage/maria/ma_blockrec.h: in-write hooks move from ma_loghandler.c storage/maria/ma_check.c: more straightforward size of buffer storage/maria/ma_checkpoint.c: <= is enough storage/maria/ma_commit.c: new prototype of translog_write_record() storage/maria/ma_create.c: new prototype of translog_write_record() storage/maria/ma_delete.c: The row's checksum must be computed before calling(*delete_record)(), not after, because it must be known inside _ma_delete_block_record() (to update the table's live checksum when writing UNDO_ROW_DELETE). If deleting from a transactional table, live checksum was already updated when writing UNDO_ROW_DELETE. storage/maria/ma_delete_all.c: @todo is now done (in ma_loghandler.c) storage/maria/ma_delete_table.c: new prototype of translog_write_record() storage/maria/ma_loghandler.c: * in-write hooks move to ma_blockrec.c. * translog_write_record() gets a new argument 'hook_arg' which is passed down to pre|inwrite_rec_hook. It is more useful that 'parts' for those hooks, because when those hooks are called, 'parts' has possibly been mangled (like with LSN compression) and is so unpredictable. * fix for compiler warning (unused buffer_start when compiling without debug support) * Because checksum delta is stored into UNDO_ROW_INSERT|UPDATE|DELETE and CLR_END, but only if the table has live checksum, these records are not PSEUDOFIXEDLENGTH anymore, they are now VARIABLE_LENGTH (their length is X if no live checksum and X+4 otherwise). * add an inwrite_rec_hook for UNDO_ROW_UPDATE, which updates the table's live checksum. Update it also in hooks of UNDO_ROW_INSERT| DELETE and REDO_DELETE_ALL and CLR_END. * Bugfix: when reading a record in translog_read_record(), it happened that "length" became negative, because the function assumed that the record extended beyond the page's end, whereas it may be shorter. storage/maria/ma_loghandler.h: * Instead of hard-coded "1" and "4", use symbols and macros to store/retrieve the type of record which the CLR_END corresponds to, and the checksum variation caused by the operation which logs the record * translog_write_record() gets a new argument 'hook_arg' which is passed down to pre|inwrite_rec_hook. It is more useful that 'parts' for those hooks, because when those hooks are called, 'parts' has possibly been mangled (like with LSN compression) and is so unpredictable. storage/maria/ma_open.c: fix for "empty body in if() statement" (when compiling without safemutex) storage/maria/ma_pagecache.c: <= is enough storage/maria/ma_recovery.c: * print the time that each recovery phase (REDO/UNDO/flush) took; this is enabled only when recovering from ha_maria. Is it printed n seconds with a fractional part of one digit (like 123.4 seconds). * In the REDO phase, update the table's live checksum by using the checksum delta stored in UNDO_ROW_INSERT|DELETE|UPDATE and CLR_END. Update it too when seeing REDO_DELETE_ALL. * In the UNDO phase, when executing UNDO_ROW_INSERT, if the table does not have live checksum then reading the record's header (as done by the master loop of run_undo_phase()) is enough; otherwise we do a translog_read_record() to have the checksum delta ready for _ma_apply_undo_row_insert(). * When at the end of the REDO phase we notice that there is an unfinished group of REDOs, don't assert in debug binaries, as I verified that it can happen in real life (with kill -9) * removing ' in #error as it confuses gcc3 storage/maria/ma_rename.c: new prototype of translog_write_record() storage/maria/ma_test_recovery.expected: Change in output of ma_test_recovery: now all live checksums of original tables equal those of tables recreated by the REDO phase and those of tables fixed by the UNDO phase. I.e. recovery of the live checksum looks like working (which was after all the only goal of this changeset). I checked by hand that it's not just all live checksums which are now 0 and that's why they match. They are the old values like 3757530372. maria.test has hard-coded checksum values in its result file so checks this too. storage/maria/ma_update.c: * It's useless to put up HA_STATE_CHANGED in 'key_changed', as we put up HA_STATE_CHANGED in info->update anyway. * We need to compute the old and new rows' checksum before calling (*update_record)(), as checksum delta must be known when logging UNDO_ROW_UPDATE which is done by _ma_update_block_record(). Note that some functions change the 'newrec' record (at least _ma_check_unique() does) so we cannot move the checksum computation too early in the function. storage/maria/ma_write.c: If inserting into a transactional table, live's checksum was already updated when writing UNDO_ROW_INSERT. The multiplication is a trick to save an if(). storage/maria/unittest/ma_test_loghandler-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_first_lsn-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_max_lsn-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_multigroup-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_multithread-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_noflush-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_pagecache-t.c: new prototype of translog_write_record() storage/maria/unittest/ma_test_loghandler_purge-t.c: new prototype of translog_write_record() storage/myisam/sort.c: fix for compiler warnings in pushbuild (write_merge_key* functions didn't have their declaration match MARIA_HA::write_key).
2007-10-02 18:02:09 +02:00
compile_time_assert(sizeof(int) <= 4);
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
pthread_mutex_lock(&share->intern_lock);
/*
Tables in a normal state have their two file descriptors open.
In some rare cases like REPAIR, some descriptor may be closed or even
-1. If that happened, the _ma_state_info_write() may fail. This is
prevented by enclosing all all places which close/change kfile.file with
intern_lock.
*/
kfile= share->kfile;
dfile= share->bitmap.file;
/*
Ignore table which has no logged writes (all its future log records will
be found naturally by Recovery). Ignore obsolete shares (_before_
setting themselves to last_version=0 they already did all flush and
sync; if we flush their state now we may be flushing an obsolete state
onto a newer one (assuming the table has been reopened with a different
share but of course same physical index file).
*/
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
ignore_share= (share->id == 0) | (share->last_version == 0);
DBUG_PRINT("info", ("ignore_share: %d", ignore_share));
if (!ignore_share)
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
{
/** @todo avoid strlen */
uint open_file_name_len= strlen(share->open_file_name) + 1;
/* remember the descriptors for background flush */
*(dfiles_end++)= dfile;
*(kfiles_end++)= kfile;
/* we will store this table in the record */
nb_stored++;
int2store(ptr, share->id);
ptr+= 2;
lsn_store(ptr, share->lsn_of_file_id);
ptr+= LSN_STORE_SIZE;
/*
first_bitmap_with_space is not updated under log's lock, and is
important. We would need the bitmap's lock to get it right. Recovery
of this is not clear, so we just play safe: write it out as
unknown: if crash, _ma_bitmap_init() at next open (for example in
Recovery) will convert it to 0 and thus the first insertion will
search for free space from the file's first bitmap (0) -
under-optimal but safe.
If no crash, maria_close() will write the exact value.
*/
state_copy->state.first_bitmap_with_space= ~(ulonglong)0;
memcpy(ptr, share->open_file_name, open_file_name_len);
ptr+= open_file_name_len;
if (cmp_translog_addr(share->state.is_of_horizon,
checkpoint_start_log_horizon) >= 0)
{
/*
State was flushed recently, it does not hold down the log's
low-water mark and will not give avoidable work to Recovery. So we
needn't flush it. Also, it is possible that while we copied the
state above (under log's lock, without intern_lock) it was being
modified in memory or flushed to disk (without log's lock, under
intern_lock, like in maria_extra()), so our copy may be incorrect
and we should not flush it.
It may also be a share which got last_version==0 since we checked
last_version; in this case, it flushed its state and the LSN test
above will catch it.
*/
}
else
{
/*
We could do the state flush only if share->changed, but it's
tricky.
Consider a maria_write() which has written REDO,UNDO, and before it
calls _ma_writeinfo() (setting share->changed=1), checkpoint
happens and sees share->changed=0, does not flush state. It is
possible that Recovery does not start from before the REDO and thus
the state is not recovered. A solution may be to set
share->changed=1 under log mutex when writing log records.
But as anyway we have another problem below, this optimization would
be of little use.
*/
/** @todo flush state only if changed since last checkpoint */
DBUG_ASSERT(share->last_version != 0);
state_copy->state.is_of_horizon= share->state.is_of_horizon=
state_copies_horizon;
if (kfile.file >= 0)
sync_error|=
_ma_state_info_write_sub(kfile.file, &state_copy->state, 1);
/*
We don't set share->changed=0 because it may interfere with a
concurrent _ma_writeinfo() doing share->changed=1 (cancel its
effect). The sad consequence is that we will flush the same state at
each checkpoint if the table was once written and then not anymore.
*/
}
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
if (_ma_bitmap_flush_all(share))
{
sync_error= 1;
/** @todo all write failures should mark table corrupted */
ma_message_no_user(0, "checkpoint bitmap page flush failed");
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
DBUG_ASSERT(share->pagecache == maria_pagecache);
}
if (share->in_checkpoint & MARIA_CHECKPOINT_SHOULD_FREE_ME)
{
/* maria_close() left us to free the share */
pthread_mutex_unlock(&share->intern_lock);
pthread_mutex_destroy(&share->intern_lock);
my_free((uchar *)share, MYF(0));
}
else
{
/* share goes back to normal state */
share->in_checkpoint= 0;
pthread_mutex_unlock(&share->intern_lock);
}
/*
We do the big disk writes out of intern_lock to not block other
users of this table (intern_lock is taken at the start and end of
every statement). This means that file descriptors may be invalid
(files may have been closed for example by HA_EXTRA_PREPARE_FOR_*
under Windows, or REPAIR). This should not be a problem as we use
MY_IGNORE_BADFD. Descriptors may even point to other files but then
the old blocks (of before the close) must have been flushed for sure,
so our flush will flush new blocks (of after the latest open) and that
should do no harm.
*/
/*
If CHECKPOINT_MEDIUM, this big flush below may result in a
serious write burst. Realize that all pages dirtied between the
last checkpoint and the one we are doing now, will be flushed at
next checkpoint, except those evicted by LRU eviction (depending on
the size of the page cache compared to the size of the working data
set, eviction may be rare or frequent).
We avoid that burst by anticipating: those pages are flushed
in bunches spanned regularly over the time interval between now and
the next checkpoint, by a background thread. Thus the next checkpoint
will have only little flushing to do (CHECKPOINT_MEDIUM should thus be
only a little slower than CHECKPOINT_INDIRECT).
*/
WL#3071 - Maria checkpoint - serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
2007-10-19 14:15:13 +02:00
/*
PageCacheFlushConcurrencyBugs
Inside the page cache, calls to flush_pagecache_blocks_int() on the same
file are serialized. Examples of concurrency bugs which happened when we
didn't have this serialization:
- maria_chk_size() (via CHECK TABLE) happens concurrently with
Checkpoint: Checkpoint is flushing a page: it pins the page and is
pre-empted, maria_chk_size() wants to flush this page too so gets an
error because Checkpoint pinned this page. Such error makes
maria_chk_size() mark the table as corrupted.
- maria_close() happens concurrently with Checkpoint:
Checkpoint is flushing a page: it registers a request on the page, is
pre-empted ; maria_close() flushes this page too with FLUSH_RELEASE:
FLUSH_RELEASE will cause a free_block() which assumes the page is in the
LRU, but it is not (as Checkpoint registered a request). Crash.
- one thread is evicting a page of the file out of the LRU: it marks it
iPC_BLOCK_IN_SWITCH and is pre-empted. Then two other threads do flushes
of the same file concurrently (like above). Then one flusher sees the
page is in switch, removes it from changed_blocks[] and puts it in its
first_in_switch, so the other flusher will not see the page at all and
return too early. If it's maria_close() which returns too early, then
maria_close() may close the file descriptor, and the other flusher, and
the evicter will fail to write their page: corruption.
*/
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
if (!ignore_share)
{
if (filter != NULL)
{
if ((flush_pagecache_blocks_with_filter(maria_pagecache,
&dfile, FLUSH_KEEP_LAZY,
filter, &filter_param) &
PCFLUSH_ERROR))
ma_message_no_user(0, "checkpoint data page flush failed");
if ((flush_pagecache_blocks_with_filter(maria_pagecache,
&kfile, FLUSH_KEEP_LAZY,
filter, &filter_param) &
PCFLUSH_ERROR))
ma_message_no_user(0, "checkpoint index page flush failed");
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/*
fsyncs the fd, that's the loooong operation (e.g. max 150 fsync
per second, so if you have touched 1000 files it's 7 seconds).
*/
WL#3072 - Maria recovery. * fix for bitmap vs checkpoint bug which could lead to corrupted tables in case of crashes at certain moments: a bitmap could be flushed to disk even though it was inconsistent with the log (it could be flushed before REDO-UNDO are written to the log). One bug remains, need code from others. Tests added. Fix is to pin unflushable bitmap pages, and let checkpoint wait for them to be flushable. * fix for long_trid!=0 assertion failure at Recovery. * less useless wakeups in the background flush|checkpoint thread. * store global_trid_generator in checkpoint record. mysql-test/r/maria-recovery.result: result update mysql-test/t/maria-recovery.test: make it easier to locate subtests storage/maria/ma_bitmap.c: When we send a bitmap to the pagecache, if this bitmap is not in a flushable state we keep it pinned and add it to a list, it will be unpinned when the bitmap is flushable again. A new function _ma_bitmap_flush_all() used by checkpoint. A new function _ma_bitmap_flushable() used by block format to signal when it starts modifying a bitmap and when it is done with it. storage/maria/ma_blockrec.c: When starting a row operation (insert/update/delete), mark that the bitmap is not flushable (because for example INSERT is going to over-allocate in the bitmap to prevent other threads from using our data pages). If a checkpoint comes at this moment it will wait for the bitmap to be flushable before flushing it. When the operation ends, bitmap becomes flushable again; that transition is done under the bitmap's mutex (needed for correct synchro with a concurrent checkpoint); but for INSERT/UPDATE this happens inside _ma_bitmap_release_unused() at a place where it already has the mutex, so the only penalty (mutex adding) is in DELETE and UNDO of INSERT. In case of errors after setting the bitmap unflushable, we must always set it back to flushable or checkpoint would block. Debug possibilities to force a sleep while the bitmap is over-allocated. In case of error in get_head_or_tail() in allocate_and_write_block_record(), we still need to unpin all pages. Bugfix: _ma_apply_redo_insert_row_blobs() produced wrong data_file_length. storage/maria/ma_blockrec.h: new bitmap calls. storage/maria/ma_checkpoint.c: filter_flush_indirect not needed anymore (flushing bitmap pages happens in _ma_bitmap_flush_all() now). So st_filter_param::is_data_file|pages_covered_by_bitmap not needed. Other filter_flush* don't need to flush bitmap anymore. Add debug possibility to flush all bitmap pages outside of a checkpoint, to simulate pagecache LRU eviction. When the background flush/checkpoint thread notices it has nothing to flush, it now sleeps directly until the next potential checkpoint moment instead of waking up every second. When in checkpoint we decide to not store a table in the checkpoint record (because it has logged no writes for example), we can also skip flushing this table. storage/maria/ma_commit.c: comment is out-of-date storage/maria/ma_key_recover.c: comment fix storage/maria/ma_loghandler.c: comment is out-of-date storage/maria/ma_open.c: comment is out-of-date storage/maria/ma_pagecache.c: comment for bug to fix. And we don't take checkpoints at end of REDO phase yet so can trust block->type. storage/maria/ma_recovery.c: Comments. Now-unneeded code for incomplete REDO-UNDO groups removed. When we forget about an old transaction we must really forget about it with bzero() (fixes the "long_trid!=0 assertion" recovery bug). When we delete a row with maria_delete() we turn on STATE_NOT_OPTIMIZED_ROWS so we do the same when we see a CLR_END for an UNDO_ROW_INSERT or when we execute an UNDO_ROW_INSERT (in both cases a row was deleted). Pick up max_long_trid from the checkpoint record. storage/maria/maria_chk.c: comment storage/maria/maria_def.h: MARIA_FILE_BITMAP gets new members: 'flushable', 'bitmap_cond' and 'pinned_pages'. storage/maria/trnman.c: I used to think that recovery only needs to know the maximum TrID of the lists of active and committed transactions. But no, sometimes both lists can even be empty and their TrID should not be reused. So Checkpoint now saves global_trid_generator in the checkpoint record. storage/maria/trnman_public.h: macros to read/store a TrID mysql-test/r/maria-recovery-bitmap.result: result is ok. Without the code fix, we would get a corruption message about the bitmap page in CHECK TABLE EXTENDED. mysql-test/t/maria-recovery-bitmap-master.opt: usual when we crash mysqld in tests mysql-test/t/maria-recovery-bitmap.test: test of recovery problems specific of the bitmap pages.
2007-12-14 16:14:12 +01:00
sync_error|=
my_sync(dfile.file, MYF(MY_WME | MY_IGNORE_BADFD)) |
my_sync(kfile.file, MYF(MY_WME | MY_IGNORE_BADFD));
/*
in case of error, we continue because writing other tables to disk is
still useful.
*/
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
}
if (sync_error)
goto err;
/* We maybe over-estimated (due to share->id==0 or last_version==0) */
DBUG_ASSERT(str->length >= (uint)(ptr - str->str));
str->length= (uint)(ptr - str->str);
/*
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
As we support max 65k tables open at a time (2-byte short id), we
assume uint is enough for the cumulated length of table names; and
LEX_STRING::length is uint.
*/
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
int4store(str->str, nb_stored);
error= unmark_tables= 0;
err:
if (unlikely(unmark_tables))
{
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
/* maria_close() uses THR_LOCK_maria from start to end */
pthread_mutex_lock(&THR_LOCK_maria);
for (i= 0; i < nb; i++)
{
MARIA_SHARE *share= distinct_shares[i];
if (share->in_checkpoint & MARIA_CHECKPOINT_SHOULD_FREE_ME)
{
/* maria_close() left us to free the share */
pthread_mutex_destroy(&share->intern_lock);
my_free((uchar *)share, MYF(0));
}
else
{
/* share goes back to normal state */
share->in_checkpoint= 0;
}
}
pthread_mutex_unlock(&THR_LOCK_maria);
}
WL#3071 Maria checkpoint Finally this is the real checkpoint code. It however exhibits unstabilities when a checkpoint runs concurrently with data-modifying clients (table corruption, transaction log's assertions) so for now a checkpoint is taken only at startup after recovery and at shutdown, i.e. not in concurrent situations. Later we will let it run periodically, as well as flush dirty pages periodically (almost all needed code is there already, only pagecache code is written but not committed). WL#3072 Maria recovery * replacing UNDO_ROW_PURGE with CLR_END; testing of those CLR_END via ma_test2 which has INSERTs failing with duplicate keys. * replaying of REDO_RENAME_TABLE Now, off to test Recovery in ha_maria :) BitKeeper/deleted/.del-ma_least_recently_dirtied.c: Delete: storage/maria/ma_least_recently_dirtied.c BitKeeper/deleted/.del-ma_least_recently_dirtied.h: Delete: storage/maria/ma_least_recently_dirtied.h storage/maria/Makefile.am: compile Checkpoint module storage/maria/ha_maria.cc: When ha_maria starts, do a recovery from last checkpoint. Take a checkpoint when that recovery has ended and when ha_maria shuts down cleanly. storage/maria/ma_blockrec.c: * even if my_sync() fails we have to my_close() (otherwise we leak a descriptor) * UNDO_ROW_PURGE is replaced by a simple CLR_END for UNDO_ROW_INSERT, as promised in the old comment; it gives us skipping during the UNDO phase. storage/maria/ma_check.c: All REDOs before create_rename_lsn are ignored by Recovery. So create_rename_lsn must be set only after all data/index has been flushed and forced to disk. We thus move write_log_record_for_repair() to after _ma_flush_tables_files_after_repair(). storage/maria/ma_checkpoint.c: Checkpoint module. storage/maria/ma_checkpoint.h: optional argument if caller wants a thread to periodically take checkpoints and flush dirty pages. storage/maria/ma_create.c: * no need to init some vars as the initial bzero(share) takes care of this. * update to new function's name * even if we fail in my_sync() we have to my_close() storage/maria/ma_extra.c: Checkpoint reads share->last_version under intern_lock, so we make maria_extra() update it under intern_lock. THR_LOCK_maria still needed because of _ma_test_if_reopen(). storage/maria/ma_init.c: destroy checkpoint module when Maria shuts down. storage/maria/ma_loghandler.c: * UNDO_ROW_PURGE gone (see ma_blockrec.c) * we need to remember the LSN of the LOGREC_FILE_ID for a share, because this LSN is needed into the checkpoint record (Recovery wants to know the validity domain of an id->name mapping) * translog_get_horizon_no_lock() needed for Checkpoint * comment about failing assertion (Sanja knows) * translog_init_reader_data() thought that translog_read_record_header_scan() returns 0 in case of error, but 0 just means "0-length header". * translog_assign_id_to_share() now needs the MARIA_HA because LOGREC_FILE_ID uses a log-write hook. * Verify that (de)assignment of share->id happens only under intern_lock, as Checkpoint reads this id with intern_lock. * translog_purge() can accept TRANSLOG_ADDRESS, not necessarily a real LSN. storage/maria/ma_loghandler.h: prototype updates storage/maria/ma_open.c: no need to initialize "res" storage/maria/ma_pagecache.c: When taking a checkpoint, we don't need to know the maximum rec_lsn of dirty pages; this LSN was intended to be used in the two-checkpoint rule, but last_checkpoint_lsn is as good. 4 bytes for stored_list_size is enough as PAGECACHE::blocks (number of blocks which the pagecache can contain) is int. storage/maria/ma_pagecache.h: new prototype storage/maria/ma_recovery.c: * added replaying of REDO_RENAME_TABLE * UNDO_ROW_PURGE gone (see ma_blockrec.c), replaced by CLR_END * Recovery from the last checkpoint record now possible * In new_table() we skip the table if the id->name mapping is older than create_rename_lsn (mapping dates from lsn_of_file_id). * in get_MARIA_HA_from_REDO_record() we skip the record if the id->name mapping is newer than the record (can happen if processing a record which is before the checkpoint record). * parse_checkpoint_record() has to return a LSN, that's what caller expects storage/maria/ma_rename.c: new function's name; log end zeroes of tables' names (ease recovery) storage/maria/ma_test2.c: * equivalent of ma_test1's --test-undo added (named -u here). * -t=1 now stops right after creating the table, so that we can test undoing of INSERTs with duplicate keys (which tests the CLR_END logged by _ma_write_abort_block_record()). storage/maria/ma_test_recovery.expected: Result of testing undoing of INSERTs with duplicate keys; there are some differences in maria_chk -dvv but they are normal (removing records does not shrink data/index file, does not put back the "analyzed, optimized keys"(etc) index state. storage/maria/ma_test_recovery: Test undoing of INSERTs with duplicate keys, using ma_test2; when such INSERT happens, it logs REDO_INSERT, UNDO_INSERT, REDO_DELETE, CLR_END; we abort after that, and test that CLR_END causes recovery to jump over UNDO_INSERT. storage/maria/ma_write.c: comment storage/maria/maria_chk.c: comment storage/maria/maria_def.h: * a new bit in MARIA_SHARE::in_checkpoint, used to build a list of unique shares during Checkpoint. * MARIA_SHARE::lsn_of_file_id added: the LSN of the last LOGREC_FILE_ID for this share; needed to know to which LSN domain the mappings found in the Checkpoint record apply (new mappings should not apply to old REDOs). storage/maria/trnman.c: * small changes to how trnman_collect_transactions() fills its buffer; it also uses a non-dummy lsn_read_non_atomic() found in ma_checkpoint.h
2007-09-12 11:27:34 +02:00
my_free((uchar *)distinct_shares, MYF(MY_ALLOW_ZERO_PTR));
my_free((uchar *)state_copies, MYF(MY_ALLOW_ZERO_PTR));
DBUG_RETURN(error);
}