mariadb/storage/maria/trnman.c
unknown 5099033c26 WL#3138: Maria - fast "SELECT COUNT(*) FROM t;" and "CHECKSUM TABLE t"
Added argument to maria_end_bulk_insert() to know if the table will be deleted after the operation
Fixed wrong call to strmake
Don't call bulk insert in case of inserting only one row (speed optimization as starting/stopping bulk insert
Allow storing year 2155 in year field
When running with purify/valgrind avoid copying structures over themself
Added hook 'trnnam_end_trans_hook' that is called when transaction ends
Added trn->used_tables that is used to an entry for all tables used by transaction
Fixed that ndb doesn't crash on duplicate key error when start_bulk_insert/end_bulk_insert are not called


include/maria.h:
  Added argument to maria_end_bulk_insert() to know if the table will be deleted after the operation
include/my_tree.h:
  Added macro 'reset_free_element()' to be able to ignore calls to the external free function.
  Is used to optimize end-bulk-insert in case of failures, in which case we don't want write the remaining keys in the tree
mysql-test/install_test_db.sh:
  Upgrade to new mysql_install_db options
mysql-test/r/maria-mvcc.result:
  New tests
mysql-test/r/maria.result:
  New tests
mysql-test/suite/ndb/r/ndb_auto_increment.result:
  Fixed error message now when bulk insert is not always called
mysql-test/suite/ndb/t/ndb_auto_increment.test:
  Fixed error message now when bulk insert is not always called
mysql-test/t/maria-mvcc.test:
  Added testing of versioning of count(*)
mysql-test/t/maria-page-checksum.test:
  Added comment
mysql-test/t/maria.test:
  More tests
mysys/hash.c:
  Code style change
sql/field.cc:
  Allow storing year 2155 in year field
sql/ha_ndbcluster.cc:
  Added new argument to end_bulk_insert() to signal if the bulk insert should ignored
sql/ha_ndbcluster.h:
  Added new argument to end_bulk_insert() to signal if the bulk insert should ignored
sql/ha_partition.cc:
  Added new argument to end_bulk_insert() to signal if the bulk insert should ignored
sql/ha_partition.h:
  Added new argument to end_bulk_insert() to signal if the bulk insert should ignored
sql/handler.cc:
  Don't call get_dup_key() if there is no table object. This can happen if the handler generates a duplicate key error on commit
sql/handler.h:
  Added new argument to end_bulk_insert() to signal if the bulk insert should ignored (ie, the table will be deleted)
sql/item.cc:
  Style fix
  Removed compiler warning
sql/log_event.cc:
  Added new argument to ha_end_bulk_insert()
sql/log_event_old.cc:
  Added new argument to ha_end_bulk_insert()
sql/mysqld.cc:
  Removed compiler warning
sql/protocol.cc:
  Added DBUG
sql/sql_class.cc:
  Added DBUG
  Fixed wrong call to strmake
sql/sql_insert.cc:
  Don't call bulk insert in case of inserting only one row (speed optimization as starting/stopping bulk insert involves a lot of if's)
  Added new argument to ha_end_bulk_insert()
sql/sql_load.cc:
  Added new argument to ha_end_bulk_insert()
sql/sql_parse.cc:
  Style fixes
  Avoid goto in common senario
sql/sql_select.cc:
  When running with purify/valgrind avoid copying structures over themself.  This is not a real bug in itself, but it's a waste of cycles and causes valgrind warnings
sql/sql_select.h:
  Avoid copying structures over themself.  This is not a real bug in itself, but it's a waste of cycles and causes valgrind warnings
sql/sql_table.cc:
  Call HA_EXTRA_PREPARE_FOR_DROP if table created by ALTER TABLE is going to be dropped
  Added new argument to ha_end_bulk_insert()
storage/archive/ha_archive.cc:
  Added new argument to end_bulk_insert()
storage/archive/ha_archive.h:
  Added new argument to end_bulk_insert()
storage/federated/ha_federated.cc:
  Added new argument to end_bulk_insert()
storage/federated/ha_federated.h:
  Added new argument to end_bulk_insert()
storage/maria/Makefile.am:
  Added ma_state.c and ma_state.h
storage/maria/ha_maria.cc:
  Versioning of count(*) and checksum
  - share->state.state is now assumed to be correct, not handler->state
  - Call _ma_setup_live_state() in external lock to get count(*)/checksum versioning. In case of
    not versioned and not concurrent insertable table, file->s->state.state contains the correct state information
  
  Other things:
  - file->s -> share
  - Added DBUG_ASSERT() for unlikely case
  - Optimized end_bulk_insert() to not write anything if table is going to be deleted (as in failed alter table)
  - Indentation changes in external_lock becasue of removed 'goto' caused a big conflict even if very little was changed
storage/maria/ha_maria.h:
  New argument to end_bulk_insert()
storage/maria/ma_blockrec.c:
  Update for versioning of count(*) and checksum
  Keep share->state.state.data_file_length up to date (not info->state->data_file_length)
  Moved _ma_block_xxxx_status() and maria_versioning() functions to ma_state.c
storage/maria/ma_check.c:
  Update and use share->state.state instead of info->state
  info->s to share
  Update info->state at end of repair
  Call _ma_reset_state() to update share->state_history at end of repair
storage/maria/ma_checkpoint.c:
  Call _ma_remove_not_visible_states() on checkpoint to clean up not visible state history from tables
storage/maria/ma_close.c:
  Remember state history for running transaction even if table is closed
storage/maria/ma_commit.c:
  Ensure we always call trnman_commit_trn() even if other calls fails. If we don't do that, the translog and state structures will not be freed
storage/maria/ma_delete.c:
  Versioning of count(*) and checksum:
  - Always update info->state->checksum and info->state->records
storage/maria/ma_delete_all.c:
  Versioning of count(*) and checksum:
  - Ensure that share->state.state is updated, as here is where we store the primary information
storage/maria/ma_dynrec.c:
  Use lock_key_trees instead of concurrent_insert to check if trees should be locked.
  This allows us to lock trees both for concurrent_insert and for index versioning.
storage/maria/ma_extra.c:
  Versioning of count(*) and checksum:
  - Use share->state.state instead of info->state
  - share->concurrent_insert -> share->non_transactional_concurrent_insert
  - Don't update share->state.state from info->state if transactional table
  
  Optimization:
  - Don't flush io_cache or bitmap if we are using FLUSH_IGNORE_CHANGED
storage/maria/ma_info.c:
  Get most state information from current state
storage/maria/ma_init.c:
  Add hash table and free function to store states for closed tables
  Install hook for transaction commit/rollback to update history state
storage/maria/ma_key_recover.c:
  Versioning of count(*) and checksum:
  - Use share->state.state instead of info->state
storage/maria/ma_locking.c:
  Versioning of count(*) and checksum:
  - Call virtual functions (if exists) to restore/update status
  - Move _ma_xxx_status() functions to ma_state.c
  
  info->s -> share
storage/maria/ma_open.c:
  Versioning of count(*) and checksum:
  - For not transactional tables, set info->state to point to new allocated state structure.
  - Initialize new info->state_start variable that points to state at start of transaction
  - Copy old history states from hash table (maria_stored_states) first time the table is opened
  - Split flag share->concurrent_insert to non_transactional_concurrent_insert & lock_key_tree
  - For now, only enable versioning of tables without keys (to be fixed in soon!)
  - Added new virtual function to restore status in maria_lock_database)
  
  More DBUG
storage/maria/ma_page.c:
  Versioning of count(*) and checksum:
  - Use share->state.state instead of info->state
  - Modify share->state.state.key_file_length under share->intern_lock
storage/maria/ma_range.c:
  Versioning of count(*) and checksum:
  - Lock trees based on share->lock_key_trees
  
  info->s -> share
storage/maria/ma_recovery.c:
  Versioning of count(*) and checksum:
  - Use share->state.state instead of info->state
  - Update state information on close and when reenabling logging
storage/maria/ma_rkey.c:
  Versioning of count(*) and checksum:
  - Lock trees based on share->lock_key_trees
storage/maria/ma_rnext.c:
  Versioning of count(*) and checksum:
  - Lock trees based on share->lock_key_trees
storage/maria/ma_rnext_same.c:
  Versioning of count(*) and checksum:
  - Lock trees based on share->lock_key_trees
  - Only skip rows based on file length if non_transactional_concurrent_insert is set
storage/maria/ma_rprev.c:
  Versioning of count(*) and checksum:
  - Lock trees based on share->lock_key_trees
storage/maria/ma_rsame.c:
  Versioning of count(*) and checksum:
  - Lock trees based on share->lock_key_trees
storage/maria/ma_sort.c:
  Use share->state.state instead of info->state
  Fixed indentation
storage/maria/ma_static.c:
  Added maria_stored_state
storage/maria/ma_update.c:
  Versioning of count(*) and checksum:
  - Always update info->state->checksum and info->state->records
  - Remove optimization for index file update as it doesn't work for transactional tables
storage/maria/ma_write.c:
  Versioning of count(*) and checksum:
  - Always update info->state->checksum and info->state->records
storage/maria/maria_def.h:
  Move MARIA_STATUS_INFO to ma_state.h
  
  Changes to MARIA_SHARE:
  - Added state_history to store count(*)/checksum states
  - Added in_trans as counter if table is used by running transactions
  - Split concurrent_insert into lock_key_trees and on_transactional_concurrent_insert.
  - Added virtual function lock_restore_status
  
  Changes to MARIA_HA:
  - save_state -> state_save
  - Added state_start to store state at start of transaction
storage/maria/maria_pack.c:
  Versioning of count(*) and checksum:
  - Use share->state.state instead of info->state
  
  Indentation fixes
storage/maria/trnman.c:
  Added hook 'trnnam_end_trans_hook' that is called when transaction ends
  Added trn->used_tables that is used to an entry for all tables used by transaction
  More DBUG
  Changed return type of trnman_end_trn() to my_bool
  Added trnman_get_min_trid() to get minimum trid in use.
  Added trnman_exists_active_transactions() to check if there exist a running transaction started between two commit id
storage/maria/trnman.h:
  Added 'used_tables'
  Moved all pointers into same groups to get better memory alignment
storage/maria/trnman_public.h:
  Added prototypes for new functions and variables
  Chagned return type of trnman_end_trn() to my_bool
storage/myisam/ha_myisam.cc:
  Added argument to end_bulk_insert() if operation should be aborted
storage/myisam/ha_myisam.h:
  Added argument to end_bulk_insert() if operation should be aborted
storage/maria/ma_state.c:
  Functions to handle state of count(*) and checksum
storage/maria/ma_state.h:
  Structures and declarations to handle state of count(*) and checksum
2008-05-29 18:33:33 +03:00

854 lines
25 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= ~(ulong) 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= ~(ulong) 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.
*/
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 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;
}