mariadb/storage/maria/trnman.c
Michael Widenius 52cb0c24a6 Added versioning of Maria index
Store max_trid in index file as state.create_trid. This is used to pack all transids in the index pages relative to max possible transid for file.
Enable versioning for transactional tables with index. Tables with an auto-increment key, rtree or fulltext keys are not versioned.
Changed info->lastkey to type MARIA_KEY. Removed info->lastkey_length as this is now part of info->lastkey
Renamed old info->lastkey to info->lastkey_buff
Use exact key lenghts for keys, not USE_WHOLE_KEY
For partial key searches, use SEARCH_PART_KEY
When searching to insert new key on page, use SEARCH_INSERT to mark that key has rowid

Changes done in a lot of files:
- Modified functions to use MARIA_KEY instead of key pointer and key length
- Use keyinfo->root_lock instead of share->key_root_lock[keynr]
- Simplify code by using local variable keyinfo instead if share->keyinfo[i]
- Added #fdef EXTERNAL_LOCKING around removed state elements
- HA_MAX_KEY_BUFF -> MARIA_MAX_KEY_BUFF (to reserve space for transid)
- Changed type of 'nextflag' to uint32 to ensure all SEARCH_xxx flags fits into it

.bzrignore:
  Added missing temporary directory
extra/Makefile.am:
  comp_err is now deleted on make distclean
include/maria.h:
  Added structure MARIA_KEY, which is used for intern key objects in Maria.
  Changed functions to take MARIA_KEY as an argument instead of pointer to packed key.
  Changed some functions that always return true or false to my_bool.
  Added virtual function make_key() to avoid if in _ma_make_key()
  Moved rw_lock_t for locking trees from share->key_root_lock to MARIA_KEYDEF. This makes usage of the locks simpler and faster
include/my_base.h:
  Added HA_RTREE_INDEX flag to mark rtree index. Used for easier checks in ma_check()
  Added SEARCH_INSERT to be used when inserting new keys
  Added SEARCH_PART_KEY for partial searches
  Added SEARCH_USER_KEY_HAS_TRANSID to be used when key we use for searching in btree has a TRANSID
  Added SEARCH_PAGE_KEY_HAS_TRANSID to be used when key we found in btree has a transid
include/my_handler.h:
  Make next_flag 32 bit to make sure we can handle all SEARCH_ bits
mysql-test/include/maria_empty_logs.inc:
  Read and restore current database; Don't assume we are using mysqltest.
  Don't log use databasename to log. Using this include should not cause any result changes.
mysql-test/r/maria-gis-rtree-dynamic.result:
  Updated results after adding some check table commands to help pinpoint errors
mysql-test/r/maria-mvcc.result:
  New tests
mysql-test/r/maria-purge.result:
  New result after adding removal of logs
mysql-test/r/maria-recovery-big.result:
  maria_empty_logs doesn't log 'use mysqltest' anymore
mysql-test/r/maria-recovery-bitmap.result:
  maria_empty_logs doesn't log 'use mysqltest' anymore
mysql-test/r/maria-recovery-rtree-ft.result:
  maria_empty_logs doesn't log 'use mysqltest' anymore
mysql-test/r/maria-recovery.result:
  maria_empty_logs doesn't log 'use mysqltest' anymore
mysql-test/r/maria.result:
  New tests
mysql-test/r/variables-big.result:
  Don't log id as it's not predictable
mysql-test/suite/rpl_ndb/r/rpl_truncate_7ndb_2.result:
  Updated results to new binlog results. (Test has not been run in a long time as it requires --big)
mysql-test/suite/rpl_ndb/t/rpl_truncate_7ndb_2-master.opt:
  Moved file to ndb replication test directory
mysql-test/suite/rpl_ndb/t/rpl_truncate_7ndb_2.test:
  Fixed wrong path to included tests
mysql-test/t/maria-gis-rtree-dynamic.test:
  Added some check table commands to help pinpoint errors
mysql-test/t/maria-mvcc.test:
  New tests
mysql-test/t/maria-purge.test:
  Remove logs to make test results predictable
mysql-test/t/maria.test:
  New tests for some possible problems
mysql-test/t/variables-big.test:
  Don't log id as it's not predictable
mysys/my_handler.c:
  Updated function comment to reflect old code
  Changed nextflag to be uint32 to ensure we can have flags > 16 bit
  Changed checking if we are in insert with NULL keys as next_flag can now include additional bits that have to be ignored.
  Added SEARCH_INSERT flag to be used when inserting new keys in btree. This flag tells us the that the keys includes row position and it's thus safe to remove SEARCH_FIND
  Added comparision of transid. This is only done if the keys actually have a transid, which is indicated by nextflag
mysys/my_lock.c:
  Fixed wrong test (Found by Guilhem)
scripts/Makefile.am:
  Ensure that test programs are deleted by make clean
sql/rpl_rli.cc:
  Moved assignment order to fix compiler warning
storage/heap/hp_write.c:
  Add SEARCH_INSERT to signal ha_key_cmp that we we should also compare rowid for keys
storage/maria/Makefile.am:
  Remove also maria log files when doing make distclean
storage/maria/ha_maria.cc:
  Use 'file->start_state' as default state for transactional tables without versioning
  At table unlock, set file->state to point to live state. (Needed for information schema to pick up right number of rows)
  In ha_maria::implicit_commit() move all locked (ie open) tables to new transaction. This is needed to ensure ha_maria->info doesn't point to a deleted history event.
  Disable concurrent inserts for insert ... select and table changes with subqueries if statement based replication as this would cause wrong results on slave
storage/maria/ma_blockrec.c:
  Updated comment
storage/maria/ma_check.c:
  Compact key pages (removes transid) when doing --zerofill
  Check that 'page_flag' on key pages contains KEYPAGE_FLAG_HAS_TRANSID if there is a single key on the page with a transid
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Use new interface to _ma_rec_pos(), _ma_dpointer(), _ma_ft_del(), ma_update_state_lsn()
  Removed not needed argument from get_record_for_key()
  Fixed that we check doesn't give errors for RTREE; We now treath these like SPATIAL
  Remove some SPATIAL specific code where the virtual functions can handle this in a general manner
  Use info->lastkey_buff instead of info->lastkey
  _ma_dpos() -> _ma_row_pos_from_key()
  _ma_make_key() -> keyinfo->make_key()
  _ma_print_key() -> _ma_print_keydata()
  _ma_move_key() -> ma_copy_copy()
  Add SEARCH_INSERT to signal ha_key_cmp that we we should also compare rowid for keys
  Ensure that data on page doesn't overwrite page checksum position
  Use DBUG_DUMP_KEY instead of DBUG_DUMP
  Use exact key lengths instead of USE_WHOLE_KEY to ha_key_cmp()
  Fixed check if rowid points outside of BLOCK_RECORD data file
  Use info->lastkey_buff instead of key on stack in some safe places
  Added #fdef EXTERNAL_LOCKING around removed state elements
storage/maria/ma_close.c:
  Use keyinfo->root_lock instead of share->key_root_lock[keynr]
storage/maria/ma_create.c:
  Removed assert that is already checked in maria_init()
  Force transactinal tables to be of type BLOCK_RECORD
  Fixed wrong usage of HA_PACK_RECORD (should be HA_OPTION_PACK_RECORD)
  Mark keys that uses HA_KEY_ALG_RTREE with HA_RTREE_INDEX for easier handling of these in ma_check
  Store max_trid in index file as state.create_trid. This is used to pack all transids in the index pages relative to max possible transid for file.
storage/maria/ma_dbug.c:
  Changed _ma_print_key() to use MARIA_KEY
storage/maria/ma_delete.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  info->lastkey2-> info->lastkey_buff2
  Added SEARCH_INSERT to signal ha_key_cmp that we we should also compare rowid for keys
  Use new interface for get_key(), _ma_get_last_key() and others
  _ma_dpos() -> ma_row_pos_from_key()
  Simplify setting of prev_key in del()
  Ensure that KEYPAGE_FLAG_HAS_TRANSID is set in page_flag if key page has transid
  Treath key pages that may have a transid as if keys would be of variable length
storage/maria/ma_delete_all.c:
  Reset history state if maria_delete_all_rows() are called
  Update parameters to _ma_update_state_lsns() call
storage/maria/ma_extra.c:
  Store and restore info->lastkey
storage/maria/ma_ft_boolean_search.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_ft_nlq_search.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Use lastkey_buff2 instead of info->lastkey+info->s->base.max_key_length (same thing)
storage/maria/ma_ft_update.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_ftdefs.h:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_fulltext.h:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_init.c:
  Check if blocksize is legal
  (Moved test here from ma_open())
storage/maria/ma_key.c:
  Added functions for storing/reading of transid 
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Moved _ma_sp_make_key() out of _ma_make_key() as we now use keyinfo->make_key to create keys
  Add transid to keys if table is versioned
  Added _ma_copy_key()
storage/maria/ma_key_recover.c:
  Add logging of page_flag (holds information if there are keys with transid on page)
  Changed DBUG_PRINT("info" -> DBUG_PRINT("redo" as the redo logging can be quite extensive
  Added lots of DBUG_PRINT()
  Added support for index page operations: KEY_OP_SET_PAGEFLAG and KEY_OP_COMPACT_PAGE
storage/maria/ma_key_recover.h:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_locking.c:
  Added new arguments to _ma_update_state_lsns_sub()
storage/maria/ma_loghandler.c:
  Fixed all logging of LSN to look similar in DBUG log
  Changed if (left != 0) to if (left) as the later is used also later in the code
storage/maria/ma_loghandler.h:
  Added new index page operations
storage/maria/ma_open.c:
  Removed allocated "state_dummy" and instead use share->state.common for transactional tables that are not versioned
  This is needed to not get double increments of state.records (one in ma_write.c and on when log is written)
  Changed info->lastkey to MARIA_KEY type
  Removed resetting of MARIA_HA variables that have 0 as default value (as info is zerofilled)
  Enable versioning for transactional tables with index. Tables with an auto-increment key, rtree or fulltext keys are not versioned.
  Check on open that state.create_trid is correct
  Extend share->base.max_key_length in case of transactional table so that it can hold transid
  Removed 4.0 compatible fulltext key mode as this is not relevant for Maria
  Removed old and wrong #ifdef ENABLE_WHEN_WE_HAVE_TRANS_ROW_ID code block
  Initialize all new virtual function pointers
  Removed storing of state->unique, state->process and store state->create_trid instead
storage/maria/ma_page.c:
  Added comment to describe key page structure
  Added functions to compact key page and log the compact operation
storage/maria/ma_range.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Use SEARCH_PART_KEY indicator instead of USE_WHOLE_KEY to detect if we are doing a part key search
  Added handling of pages with transid
storage/maria/ma_recovery.c:
  Don't assert if table we opened are not transactional. This may be a table which has been changed from transactional to not transactinal
  Added new arguments to _ma_update_state_lsns()
storage/maria/ma_rename.c:
  Added new arguments to _ma_update_state_lsns()
storage/maria/ma_rkey.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Don't use USE_WHOLE_KEY, use real length of key
  Use share->row_is_visible() to test if row is visible
  Moved search_flag == HA_READ_KEY_EXACT out of 'read-next-row' loop as this only need to be tested once
  Removed test if last_used_keyseg != 0 as this is always true
storage/maria/ma_rnext.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Simplify code by using local variable keyinfo instead if share->keyinfo[i]
  Use share->row_is_visible() to test if row is visible
storage/maria/ma_rnext_same.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  lastkey2 -> lastkey_buff2
storage/maria/ma_rprev.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Simplify code by using local variable keyinfo instead if share->keyinfo[i]
  Use share->row_is_visible() to test if row is visible
storage/maria/ma_rsame.c:
  Updated comment
  Simplify code by using local variable keyinfo instead if share->keyinfo[i]
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_rsamepos.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_rt_index.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Use better variable names
  Removed not needed casts
  _ma_dpos() -> _ma_row_pos_from_key()
  Use info->last_rtree_keypos to save position to key instead of info->int_keypos
  Simplify err: condition
  Changed return type for maria_rtree_insert() to my_bool as we are only intressed in ok/fail from this function
storage/maria/ma_rt_index.h:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_rt_key.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Simplify maria_rtree_add_key by combining idenitcal code and removing added_len
storage/maria/ma_rt_key.h:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_rt_mbr.c:
  Changed type of 'nextflag' to uint32
  Added 'to' argument to RT_PAGE_MBR_XXX functions to more clearly see which variables changes value
storage/maria/ma_rt_mbr.h:
  Changed type of 'nextflag' to uint32
storage/maria/ma_rt_split.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  key_length -> key_data_length to catch possible errors
storage/maria/ma_rt_test.c:
  Fixed wrong comment
  Reset recinfo to avoid valgrind varnings
  Fixed wrong argument to create_record() that caused test to fail
storage/maria/ma_search.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Added support of keys with optional trid
  Test for SEARCH_PART_KEY instead of USE_WHOLE_KEY to detect part key reads
  _ma_dpos() -> _ma_row_pos_from_key()
  If there may be keys with transid on the page, have _ma_bin_search() call _ma_seq_search()
  Add _ma_skip_xxx() functions to quickly step over keys (faster than calling get_key() in most cases as we don't have to copy key data)
  Combine similar code at end of _ma_get_binary_pack_key()
  Removed not used function _ma_move_key()
  In _ma_search_next() don't call _ma_search() if we aren't on a nod page.
  Update info->cur_row.trid with trid for found key
  
  
  
  Removed some not needed casts
  Added _ma_trid_from_key()
  Use MARIA_SHARE instead of MARIA_HA as arguments to _ma_rec_pos(), _ma_dpointer() and _ma_xxx_keypos_to_recpos() to make functions faster and smaller
storage/maria/ma_sort.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_sp_defs.h:
  _ma_sp_make_key() now fills in and returns (MARIA_KEY *) value
storage/maria/ma_sp_key.c:
  _ma_sp_make_key() now fills in and returns (MARIA_KEY *) value
  Don't test sizeof(double), test against 8 as we are using float8store()
  Use mi_float8store() instead of doing swap of value (same thing but faster)
storage/maria/ma_state.c:
  maria_versioning() now only calls _ma_block_get_status() if table supports versioning
  Added _ma_row_visible_xxx() functions for different occasions
  When emptying history, set info->state to point to the first history event.
storage/maria/ma_state.h:
  Added _ma_row_visible_xxx() prototypes
storage/maria/ma_static.c:
  Indentation changes
storage/maria/ma_statrec.c:
  Fixed arguments to _ma_dpointer() and _ma_rec_pos()
storage/maria/ma_test1.c:
  Call init_thr_lock() if we have versioning
storage/maria/ma_test2.c:
  Call init_thr_lock() if we have versioning
storage/maria/ma_unique.c:
  Modified functions to use MARIA_KEY
storage/maria/ma_update.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
storage/maria/ma_write.c:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Simplify code by using local variable keyinfo instead if share->keyinfo[i]
  In _ma_enlarge_root(), mark in page_flag if new key has transid
  _ma_dpos() -> _ma_row_pos_from_key()
  Changed return type of _ma_ck_write_tree() to my_bool as we are only testing if result is true or not
  Moved 'reversed' to outside block as area was used later
storage/maria/maria_chk.c:
  Added error if trying to sort with HA_BINARY_PACK_KEY
  Use new interface to get_key() and _ma_dpointer()
  _ma_dpos() -> _ma_row_pos_from_key()
storage/maria/maria_def.h:
  Modified functions to use MARIA_KEY instead of key pointer and key length
  Added 'common' to MARIA_SHARE->state for storing state for transactional tables without versioning
  Added create_trid to MARIA_SHARE
  Removed not used state variables 'process' and 'unique'
  Added defines for handling TRID's in index pages
  Changed to use MARIA_SHARE instead of MARIA_HA for some functions
  Added 'have_versioning' flag if table supports versioning
  Moved key_root_lock from MARIA_SHARE to MARIA_KEYDEF
  Changed last_key to be of type MARIA_KEY. Removed lastkey_length
  lastkey -> lastkey_buff, lastkey2 -> lastkey_buff2
  Added _ma_get_used_and_nod_with_flag() for faster access to page data when page_flag is read
  Added DBUG_DUMP_KEY for easier DBUG_DUMP of a key
  Changed 'nextflag' and assocaited variables to uint32
storage/maria/maria_ftdump.c:
  lastkey -> lastkey_buff
storage/maria/trnman.c:
  Fixed wrong initialization of min_read_from and max_commit_trid
  Added trnman_get_min_safe_trid()
storage/maria/unittest/ma_test_all-t:
  Added --start-from
storage/myisam/mi_check.c:
  Added SEARCH_INSERT, as ha_key_cmp() needs it when doing key comparision for inserting key on page in rowid order
storage/myisam/mi_delete.c:
  Added SEARCH_INSERT, as ha_key_cmp() needs it when doing key comparision for inserting key on page in rowid order
storage/myisam/mi_range.c:
  Updated comment
storage/myisam/mi_write.c:
  Added SEARCH_INSERT, as ha_key_cmp() needs it when doing key comparision for inserting key on page in rowid order
storage/myisam/rt_index.c:
  Fixed wrong parameter to rtree_get_req() which could cause crash
2008-06-26 08:18:28 +03:00

880 lines
26 KiB
C

/* Copyright (C) 2006 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
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 */
#include <my_global.h>
#include <my_sys.h>
#include <m_string.h>
#include "trnman.h"
#include "ma_checkpoint.h"
#include "ma_control_file.h"
/*
status variables:
how many trns in the active list currently,
in the committed list currently, allocated since startup.
*/
uint trnman_active_transactions, trnman_committed_transactions,
trnman_allocated_transactions;
/* list of active transactions in the trid order */
static TRN active_list_min, active_list_max;
/* list of committed transactions in the trid order */
static TRN committed_list_min, committed_list_max;
/* a counter, used to generate transaction ids */
static TrID global_trid_generator;
/* the mutex for everything above */
static pthread_mutex_t LOCK_trn_list;
/* LIFO pool of unused TRN structured for reuse */
static TRN *pool;
/* a hash for committed transactions that maps trid to a TRN structure */
static LF_HASH trid_to_committed_trn;
/* an array that maps short_trid of an active transaction to a TRN structure */
static TRN **short_trid_to_active_trn;
/* locks for short_trid_to_active_trn and pool */
static my_atomic_rwlock_t LOCK_short_trid_to_trn, LOCK_pool;
static my_bool default_trnman_end_trans_hook(TRN *, my_bool, my_bool);
my_bool (*trnman_end_trans_hook)(TRN *, my_bool, my_bool)=
default_trnman_end_trans_hook;
/*
Simple interface functions
QQ: if they stay so simple, should we make them inline?
*/
uint trnman_increment_locked_tables(TRN *trn)
{
return trn->locked_tables++;
}
uint trnman_has_locked_tables(TRN *trn)
{
return trn->locked_tables;
}
uint trnman_decrement_locked_tables(TRN *trn)
{
return --trn->locked_tables;
}
void trnman_reset_locked_tables(TRN *trn, uint locked_tables)
{
trn->locked_tables= locked_tables;
}
static my_bool
default_trnman_end_trans_hook(TRN *trn __attribute__ ((unused)),
my_bool commit __attribute__ ((unused)),
my_bool active_transactions
__attribute__ ((unused)))
{
return 0;
}
/*
NOTE
Just as short_id doubles as loid, this function doubles as
short_trid_to_LOCK_OWNER. See the compile-time assert below.
*/
#ifdef NOT_USED
static TRN *short_trid_to_TRN(uint16 short_trid)
{
TRN *trn;
compile_time_assert(offsetof(TRN, locks) == 0);
my_atomic_rwlock_rdlock(&LOCK_short_trid_to_trn);
trn= my_atomic_loadptr((void **)&short_trid_to_active_trn[short_trid]);
my_atomic_rwlock_rdunlock(&LOCK_short_trid_to_trn);
return (TRN *)trn;
}
#endif
static uchar *trn_get_hash_key(const uchar *trn, size_t *len,
my_bool unused __attribute__ ((unused)))
{
*len= sizeof(TrID);
return (uchar *) & ((*((TRN **)trn))->trid);
}
/**
@brief Initializes transaction manager.
@param initial_trid Generated TrIDs will start from initial_trid+1.
@return Operation status
@retval 0 OK
@retval !=0 Error
*/
int trnman_init(TrID initial_trid)
{
DBUG_ENTER("trnman_init");
short_trid_to_active_trn= (TRN **)my_malloc(SHORT_TRID_MAX*sizeof(TRN*),
MYF(MY_WME|MY_ZEROFILL));
if (unlikely(!short_trid_to_active_trn))
DBUG_RETURN(1);
short_trid_to_active_trn--; /* min short_trid is 1 */
/*
Initialize lists.
active_list_max.min_read_from must be larger than any trid,
so that when an active list is empty we would could free
all committed list.
And committed_list_max itself can not be freed so
committed_list_max.commit_trid must not be smaller that
active_list_max.min_read_from
*/
active_list_max.trid= active_list_min.trid= 0;
active_list_max.min_read_from= ~(TrID) 0;
active_list_max.next= active_list_min.prev= 0;
active_list_max.prev= &active_list_min;
active_list_min.next= &active_list_max;
committed_list_max.commit_trid= ~(TrID) 0;
committed_list_max.next= committed_list_min.prev= 0;
committed_list_max.prev= &committed_list_min;
committed_list_min.next= &committed_list_max;
trnman_active_transactions= 0;
trnman_committed_transactions= 0;
trnman_allocated_transactions= 0;
pool= 0;
global_trid_generator= initial_trid;
lf_hash_init(&trid_to_committed_trn, sizeof(TRN*), LF_HASH_UNIQUE,
0, 0, trn_get_hash_key, 0);
DBUG_PRINT("info", ("pthread_mutex_init LOCK_trn_list"));
pthread_mutex_init(&LOCK_trn_list, MY_MUTEX_INIT_FAST);
my_atomic_rwlock_init(&LOCK_short_trid_to_trn);
my_atomic_rwlock_init(&LOCK_pool);
#ifdef NOT_USED
lockman_init(&maria_lockman, (loid_to_lo_func *)&short_trid_to_TRN, 10000);
#endif
DBUG_RETURN(0);
}
/*
NOTE
this could only be called in the "idle" state - no transaction can be
running. See asserts below.
*/
void trnman_destroy()
{
DBUG_ENTER("trnman_destroy");
if (short_trid_to_active_trn == NULL) /* trnman already destroyed */
DBUG_VOID_RETURN;
DBUG_ASSERT(trid_to_committed_trn.count == 0);
DBUG_ASSERT(trnman_active_transactions == 0);
DBUG_ASSERT(trnman_committed_transactions == 0);
DBUG_ASSERT(active_list_max.prev == &active_list_min);
DBUG_ASSERT(active_list_min.next == &active_list_max);
DBUG_ASSERT(committed_list_max.prev == &committed_list_min);
DBUG_ASSERT(committed_list_min.next == &committed_list_max);
while (pool)
{
TRN *trn= pool;
pool= pool->next;
DBUG_ASSERT(trn->locks.mutex == 0);
DBUG_ASSERT(trn->locks.cond == 0);
my_free((void *)trn, MYF(0));
}
lf_hash_destroy(&trid_to_committed_trn);
DBUG_PRINT("info", ("pthread_mutex_destroy LOCK_trn_list"));
pthread_mutex_destroy(&LOCK_trn_list);
my_atomic_rwlock_destroy(&LOCK_short_trid_to_trn);
my_atomic_rwlock_destroy(&LOCK_pool);
my_free((void *)(short_trid_to_active_trn+1), MYF(0));
short_trid_to_active_trn= NULL;
#ifdef NOT_USED
lockman_destroy(&maria_lockman);
#endif
DBUG_VOID_RETURN;
}
/*
NOTE
TrID is limited to 6 bytes. Initial value of the generator
is set by the recovery code - being read from the last checkpoint
(or 1 on a first run).
*/
static TrID new_trid()
{
DBUG_ENTER("new_trid");
DBUG_ASSERT(global_trid_generator < 0xffffffffffffLL);
DBUG_PRINT("info", ("safe_mutex_assert_owner LOCK_trn_list"));
safe_mutex_assert_owner(&LOCK_trn_list);
DBUG_RETURN(++global_trid_generator);
}
static void set_short_trid(TRN *trn)
{
int i= (int) ((global_trid_generator + (intptr)trn) * 312089 %
SHORT_TRID_MAX + 1);
for ( ; !trn->short_id ; i= 1)
{
my_atomic_rwlock_wrlock(&LOCK_short_trid_to_trn);
for ( ; i <= SHORT_TRID_MAX; i++) /* the range is [1..SHORT_TRID_MAX] */
{
void *tmp= NULL;
if (short_trid_to_active_trn[i] == NULL &&
my_atomic_casptr((void **)&short_trid_to_active_trn[i], &tmp, trn))
{
trn->short_id= i;
break;
}
}
my_atomic_rwlock_wrunlock(&LOCK_short_trid_to_trn);
}
}
/*
DESCRIPTION
start a new transaction, allocate and initialize transaction object
mutex and cond will be used for lock waits
*/
TRN *trnman_new_trn(pthread_mutex_t *mutex, pthread_cond_t *cond,
void *stack_end)
{
TRN *trn;
DBUG_ENTER("trnman_new_trn");
/*
we have a mutex, to do simple things under it - allocate a TRN,
increment trnman_active_transactions, set trn->min_read_from.
Note that all the above is fast. generating short_trid may be slow,
as it involves scanning a large array - so it's done outside of the
mutex.
*/
DBUG_PRINT("info", ("pthread_mutex_lock LOCK_trn_list"));
pthread_mutex_lock(&LOCK_trn_list);
/* Allocating a new TRN structure */
trn= pool;
/*
Popping an unused TRN from the pool
(ABA isn't possible, we're behind a mutex
*/
my_atomic_rwlock_wrlock(&LOCK_pool);
while (trn && !my_atomic_casptr((void **)&pool, (void **)&trn,
(void *)trn->next))
/* no-op */;
my_atomic_rwlock_wrunlock(&LOCK_pool);
/* Nothing in the pool ? Allocate a new one */
if (!trn)
{
/*
trn should be completely initalized at create time to allow
one to keep a known state on it.
(Like redo_lns, which is assumed to be 0 at start of row handling
and reset to zero before end of row handling)
*/
trn= (TRN *)my_malloc(sizeof(TRN), MYF(MY_WME | MY_ZEROFILL));
if (unlikely(!trn))
{
DBUG_PRINT("info", ("pthread_mutex_unlock LOCK_trn_list"));
pthread_mutex_unlock(&LOCK_trn_list);
return 0;
}
trnman_allocated_transactions++;
}
trn->pins= lf_hash_get_pins(&trid_to_committed_trn, stack_end);
if (!trn->pins)
{
trnman_free_trn(trn);
return 0;
}
trnman_active_transactions++;
trn->min_read_from= active_list_min.next->trid;
trn->trid= new_trid();
trn->short_id= 0;
trn->next= &active_list_max;
trn->prev= active_list_max.prev;
active_list_max.prev= trn->prev->next= trn;
DBUG_PRINT("info", ("pthread_mutex_unlock LOCK_trn_list"));
pthread_mutex_unlock(&LOCK_trn_list);
if (unlikely(!trn->min_read_from))
{
/*
We are the only transaction. Set min_read_from so that we can read
our own rows
*/
trn->min_read_from= trn->trid + 1;
}
trn->commit_trid= 0;
trn->rec_lsn= trn->undo_lsn= trn->first_undo_lsn= 0;
trn->used_tables= 0;
trn->locks.mutex= mutex;
trn->locks.cond= cond;
trn->locks.waiting_for= 0;
trn->locks.all_locks= 0;
#ifdef NOT_USED
trn->locks.pins= lf_alloc_get_pins(&maria_lockman.alloc);
#endif
trn->locked_tables= 0;
/*
only after the following function TRN is considered initialized,
so it must be done the last
*/
set_short_trid(trn);
DBUG_PRINT("exit", ("trn: x%lx trid: 0x%lu",
(ulong) trn, (ulong) trn->trid));
DBUG_RETURN(trn);
}
/*
remove a trn from the active list.
if necessary - move to committed list and set commit_trid
NOTE
Locks are released at the end. In particular, after placing the
transaction in commit list, and after setting commit_trid. It's
important, as commit_trid affects visibility. Locks don't affect
anything they simply delay execution of other threads - they could be
released arbitrarily late. In other words, when locks are released it
serves as a start banner for other threads, they start to run. So
everything they may need must be ready at that point.
RETURN
0 ok
1 error
*/
my_bool trnman_end_trn(TRN *trn, my_bool commit)
{
int res= 1;
TRN *free_me= 0;
LF_PINS *pins= trn->pins;
DBUG_ENTER("trnman_end_trn");
DBUG_ASSERT(trn->rec_lsn == 0);
/* if a rollback, all UNDO records should have been executed */
DBUG_ASSERT(commit || trn->undo_lsn == 0);
DBUG_PRINT("info", ("pthread_mutex_lock LOCK_trn_list"));
pthread_mutex_lock(&LOCK_trn_list);
/* remove from active list */
trn->next->prev= trn->prev;
trn->prev->next= trn->next;
/*
if trn was the oldest active transaction, now that it goes away there
may be committed transactions in the list which no active transaction
needs to bother about - clean up the committed list
*/
if (trn->prev == &active_list_min)
{
uint free_me_count;
TRN *t;
for (t= committed_list_min.next, free_me_count= 0;
t->commit_trid < active_list_min.next->min_read_from;
t= t->next, free_me_count++) /* no-op */;
DBUG_ASSERT((t != committed_list_min.next && free_me_count > 0) ||
(t == committed_list_min.next && free_me_count == 0));
/* found transactions committed before the oldest active one */
if (t != committed_list_min.next)
{
free_me= committed_list_min.next;
committed_list_min.next= t;
t->prev->next= 0;
t->prev= &committed_list_min;
trnman_committed_transactions-= free_me_count;
}
}
/*
if transaction is committed and it was not the only active transaction -
add it to the committed list (which is used for read-from relation)
*/
if (commit && active_list_min.next != &active_list_max)
{
trn->commit_trid= global_trid_generator;
trn->next= &committed_list_max;
trn->prev= committed_list_max.prev;
trnman_committed_transactions++;
res= lf_hash_insert(&trid_to_committed_trn, pins, &trn);
/*
By going on with life is res<0, we let other threads block on
our rows (because they will never see us committed in
trid_to_committed_trn) until they timeout. Though correct, this is not a
good situation:
- if connection reconnects and wants to check if its rows have been
committed, it will not be able to do that (it will just lock on them) so
connection stays permanently in doubt
- internal structures trid_to_committed_trn and committed_list are
desynchronized.
So we should take Maria down immediately, the two problems being
automatically solved at restart.
*/
DBUG_ASSERT(res <= 0);
}
if (res)
{
/*
res == 1 means the condition in the if() above was false.
res == -1 means lf_hash_insert failed
*/
trn->next= free_me;
free_me= trn;
}
else
{
committed_list_max.prev= trn->prev->next= trn;
}
if ((*trnman_end_trans_hook)(trn, commit,
active_list_min.next != &active_list_max))
res= -1;
trnman_active_transactions--;
pthread_mutex_unlock(&LOCK_trn_list);
/* the rest is done outside of a critical section */
#ifdef NOT_USED
lockman_release_locks(&maria_lockman, &trn->locks);
#endif
trn->locks.mutex= 0;
trn->locks.cond= 0;
my_atomic_rwlock_rdlock(&LOCK_short_trid_to_trn);
my_atomic_storeptr((void **)&short_trid_to_active_trn[trn->short_id], 0);
my_atomic_rwlock_rdunlock(&LOCK_short_trid_to_trn);
/*
we, under the mutex, removed going-in-free_me transactions from the
active and committed lists, thus nobody else may see them when it scans
those lists, and thus nobody may want to free them. Now we don't
need a mutex to access free_me list
*/
/* QQ: send them to the purge thread */
while (free_me)
{
TRN *t= free_me;
free_me= free_me->next;
/*
ignore OOM here. it's harmless, and there's nothing we could do, anyway
*/
(void)lf_hash_delete(&trid_to_committed_trn, pins, &t->trid, sizeof(TrID));
trnman_free_trn(t);
}
lf_hash_put_pins(pins);
#ifdef NOT_USED
lf_pinbox_put_pins(trn->locks.pins);
#endif
DBUG_RETURN(res < 0);
}
/*
free a trn (add to the pool, that is)
note - we can never really free() a TRN if there's at least one other
running transaction - see, e.g., how lock waits are implemented in
lockman.c
The same is true for other lock-free data structures too. We may need some
kind of FLUSH command to reset them all - ensuring that no transactions are
running. It may even be called automatically on checkpoints if no
transactions are running.
*/
void trnman_free_trn(TRN *trn)
{
TRN *tmp= pool;
my_atomic_rwlock_wrlock(&LOCK_pool);
do
{
/*
without this volatile cast gcc-3.4.4 moved the assignment
down after the loop at -O2
*/
*(TRN * volatile *)&(trn->next)= tmp;
} while (!my_atomic_casptr((void **)&pool, (void **)&tmp, trn));
my_atomic_rwlock_wrunlock(&LOCK_pool);
}
/*
NOTE
here we access the hash in a lock-free manner.
It's safe, a 'found' TRN can never be freed/reused before we access it.
In fact, it cannot be freed before 'trn' ends, because a 'found' TRN
can only be removed from the hash when:
found->commit_trid < ALL (trn->min_read_from)
that is, at least
found->commit_trid < trn->min_read_from
but
found->trid >= trn->min_read_from
and
found->commit_trid > found->trid
RETURN
1 can
0 cannot
-1 error (OOM)
*/
int trnman_can_read_from(TRN *trn, TrID trid)
{
TRN **found;
my_bool can;
LF_REQUIRE_PINS(3);
if (trid < trn->min_read_from)
return 1; /* Row is visible by all transactions in the system */
if (trid >= trn->trid)
{
/*
We have now two cases
trid > trn->trid, in which case the row is from a new transaction
and not visible, in which case we should return 0.
trid == trn->trid in which case the row is from the current transaction
and we should return 1
*/
return trid == trn->trid;
}
found= lf_hash_search(&trid_to_committed_trn, trn->pins, &trid, sizeof(trid));
if (found == NULL)
return 0; /* not in the hash of committed transactions = cannot read */
if (found == MY_ERRPTR)
return -1;
can= (*found)->commit_trid < trn->trid;
lf_hash_search_unpin(trn->pins);
return can;
}
/* TODO: the stubs below are waiting for savepoints to be implemented */
void trnman_new_statement(TRN *trn __attribute__ ((unused)))
{
}
void trnman_rollback_statement(TRN *trn __attribute__ ((unused)))
{
}
/**
@brief Allocates buffers and stores in them some info about transactions
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 has the intention of doing checkpoints.
@param[out] str_act pointer to where the allocated buffer,
and its size, will be put; buffer will be filled
with info about active transactions
@param[out] str_com pointer to where the allocated buffer,
and its size, will be put; buffer will be filled
with info about committed transactions
@param[out] min_first_undo_lsn pointer to where the minimum
first_undo_lsn of all transactions will be put
@return Operation status
@retval 0 OK
@retval 1 Error
*/
my_bool trnman_collect_transactions(LEX_STRING *str_act, LEX_STRING *str_com,
LSN *min_rec_lsn, LSN *min_first_undo_lsn)
{
my_bool error;
TRN *trn;
char *ptr;
uint stored_transactions= 0;
LSN minimum_rec_lsn= LSN_MAX, minimum_first_undo_lsn= LSN_MAX;
DBUG_ENTER("trnman_collect_transactions");
DBUG_ASSERT((NULL == str_act->str) && (NULL == str_com->str));
/* validate the use of read_non_atomic() in general: */
compile_time_assert((sizeof(LSN) == 8) && (sizeof(LSN_WITH_FLAGS) == 8));
pthread_mutex_lock(&LOCK_trn_list);
str_act->length= 2 + /* number of active transactions */
LSN_STORE_SIZE + /* minimum of their rec_lsn */
TRANSID_SIZE + /* current TrID generator value */
(2 + /* short id */
6 + /* long id */
LSN_STORE_SIZE + /* undo_lsn */
#ifdef MARIA_VERSIONING /* not enabled yet */
LSN_STORE_SIZE + /* undo_purge_lsn */
#endif
LSN_STORE_SIZE /* first_undo_lsn */
) * trnman_active_transactions;
str_com->length= 4 + /* number of committed transactions */
(6 + /* long id */
#ifdef MARIA_VERSIONING /* not enabled yet */
LSN_STORE_SIZE + /* undo_purge_lsn */
#endif
LSN_STORE_SIZE /* first_undo_lsn */
) * trnman_committed_transactions;
if ((NULL == (str_act->str= my_malloc(str_act->length, MYF(MY_WME)))) ||
(NULL == (str_com->str= my_malloc(str_com->length, MYF(MY_WME)))))
goto err;
/* First, the active transactions */
ptr= str_act->str + 2 + LSN_STORE_SIZE;
transid_store(ptr, global_trid_generator);
ptr+= TRANSID_SIZE;
for (trn= active_list_min.next; trn != &active_list_max; trn= trn->next)
{
/*
trns with a short trid of 0 are not even initialized, we can ignore
them. trns with undo_lsn==0 have done no writes, we can ignore them
too. XID not needed now.
*/
uint sid;
LSN rec_lsn, undo_lsn, first_undo_lsn;
if ((sid= trn->short_id) == 0)
{
/*
Not even inited, has done nothing. Or it is the
dummy_transaction_object, which does only non-transactional
immediate-sync operations (CREATE/DROP/RENAME/REPAIR TABLE), and so
can be forgotten for Checkpoint.
*/
continue;
}
/* needed for low-water mark calculation */
if (((rec_lsn= lsn_read_non_atomic(trn->rec_lsn)) > 0) &&
(cmp_translog_addr(rec_lsn, minimum_rec_lsn) < 0))
minimum_rec_lsn= rec_lsn;
/*
trn may have logged REDOs but not yet UNDO, that's why we read rec_lsn
before deciding to ignore if undo_lsn==0.
*/
if ((undo_lsn= trn->undo_lsn) == 0) /* trn can be forgotten */
continue;
stored_transactions++;
int2store(ptr, sid);
ptr+= 2;
int6store(ptr, trn->trid);
ptr+= 6;
lsn_store(ptr, undo_lsn); /* needed for rollback */
ptr+= LSN_STORE_SIZE;
/* needed for low-water mark calculation */
if (((first_undo_lsn= lsn_read_non_atomic(trn->first_undo_lsn)) > 0) &&
(cmp_translog_addr(first_undo_lsn, minimum_first_undo_lsn) < 0))
minimum_first_undo_lsn= first_undo_lsn;
lsn_store(ptr, first_undo_lsn);
ptr+= LSN_STORE_SIZE;
#ifdef MARIA_VERSIONING /* not enabled yet */
/* to know where purging should start (last delete of this trn) */
lsn_store(ptr, trn->undo_purge_lsn);
ptr+= LSN_STORE_SIZE;
#endif
/**
@todo RECOVERY: add a comment explaining why we can dirtily read some
vars, inspired by the text of "assumption 8" in WL#3072
*/
}
str_act->length= ptr - str_act->str; /* as we maybe over-estimated */
ptr= str_act->str;
DBUG_PRINT("info",("collected %u active transactions",
(uint)stored_transactions));
int2store(ptr, stored_transactions);
ptr+= 2;
/* this LSN influences how REDOs for any page can be ignored by Recovery */
lsn_store(ptr, minimum_rec_lsn);
/* one day there will also be a list of prepared transactions */
/* do the same for committed ones */
ptr= str_com->str;
int4store(ptr, trnman_committed_transactions);
ptr+= 4;
DBUG_PRINT("info",("collected %u committed transactions",
(uint)trnman_committed_transactions));
for (trn= committed_list_min.next; trn != &committed_list_max;
trn= trn->next)
{
LSN first_undo_lsn;
int6store(ptr, trn->trid);
ptr+= 6;
#ifdef MARIA_VERSIONING /* not enabled yet */
lsn_store(ptr, trn->undo_purge_lsn);
ptr+= LSN_STORE_SIZE;
#endif
first_undo_lsn= LSN_WITH_FLAGS_TO_LSN(trn->first_undo_lsn);
if (cmp_translog_addr(first_undo_lsn, minimum_first_undo_lsn) < 0)
minimum_first_undo_lsn= first_undo_lsn;
lsn_store(ptr, first_undo_lsn);
ptr+= LSN_STORE_SIZE;
}
/*
TODO: if we see there exists no transaction (active and committed) we can
tell the lock-free structures to do some freeing (my_free()).
*/
error= 0;
*min_rec_lsn= minimum_rec_lsn;
*min_first_undo_lsn= minimum_first_undo_lsn;
goto end;
err:
error= 1;
end:
pthread_mutex_unlock(&LOCK_trn_list);
DBUG_RETURN(error);
}
TRN *trnman_recreate_trn_from_recovery(uint16 shortid, TrID longid)
{
TrID old_trid_generator= global_trid_generator;
TRN *trn;
DBUG_ASSERT(maria_in_recovery && !maria_multi_threaded);
if (unlikely((trn= trnman_new_trn(NULL, NULL, NULL)) == NULL))
return NULL;
/* deallocate excessive allocations of trnman_new_trn() */
global_trid_generator= old_trid_generator;
set_if_bigger(global_trid_generator, longid);
short_trid_to_active_trn[trn->short_id]= 0;
DBUG_ASSERT(short_trid_to_active_trn[shortid] == NULL);
short_trid_to_active_trn[shortid]= trn;
trn->trid= longid;
trn->short_id= shortid;
return trn;
}
TRN *trnman_get_any_trn()
{
TRN *trn= active_list_min.next;
return (trn != &active_list_max) ? trn : NULL;
}
/**
Returns the minimum existing transaction id
@notes
This can only be called when we have at least one running transaction.
*/
TrID trnman_get_min_trid()
{
TrID min_read_from;
if (short_trid_to_active_trn == NULL)
{
/* Transaction manager not initialize; Probably called from maria_chk */
return ~(TrID) 0;
}
pthread_mutex_lock(&LOCK_trn_list);
min_read_from= active_list_min.next->min_read_from;
pthread_mutex_unlock(&LOCK_trn_list);
return min_read_from;
}
/**
Returns the minimum possible transaction id
@notes
If there is no transactions running, returns number for next running
transaction.
If one has an active transaction, the returned number will be less or
equal to this. If one is not running in a transaction one will ge the
number for the next started transaction. This is used in create table
to get a safe minimum trid to use.
*/
TrID trnman_get_min_safe_trid()
{
TrID trid;
pthread_mutex_lock(&LOCK_trn_list);
trid= min(active_list_min.next->min_read_from,
global_trid_generator);
pthread_mutex_unlock(&LOCK_trn_list);
return trid;
}
/**
Returns maximum transaction id given to a transaction so far.
*/
TrID trnman_get_max_trid()
{
TrID id;
if (short_trid_to_active_trn == NULL)
return 0;
pthread_mutex_lock(&LOCK_trn_list);
id= global_trid_generator;
pthread_mutex_unlock(&LOCK_trn_list);
return id;
}
/**
Check if there exist an active transaction between two commit_id's
@todo
Improve speed of this.
- Store transactions in tree or skip list
- Have function to copying all active transaction id's to b-tree
and use b-tree for checking states. This could be a big win
for checkpoint that will call this function for a lot of objects.
@return
0 No transaction exists
1 There is at least on active transaction in the given range
*/
my_bool trnman_exists_active_transactions(TrID min_id, TrID max_id,
my_bool trnman_is_locked)
{
TRN *trn;
my_bool ret= 0;
if (!trnman_is_locked)
pthread_mutex_lock(&LOCK_trn_list);
for (trn= active_list_min.next; trn != &active_list_max; trn= trn->next)
{
if (trn->trid > min_id && trn->trid < max_id)
{
ret= 1;
break;
}
}
if (!trnman_is_locked)
pthread_mutex_unlock(&LOCK_trn_list);
return ret;
}