diff options
Diffstat (limited to 'sql')
| -rw-r--r-- | sql/ha_ndbcluster_binlog.cc | 1 | ||||
| -rw-r--r-- | sql/handler.cc | 615 | ||||
| -rw-r--r-- | sql/handler.h | 122 | ||||
| -rw-r--r-- | sql/log.cc | 10 | ||||
| -rw-r--r-- | sql/log_event.cc | 2 | ||||
| -rw-r--r-- | sql/rpl_injector.cc | 20 | ||||
| -rw-r--r-- | sql/sp.cc | 20 | ||||
| -rw-r--r-- | sql/sp_head.cc | 1 | ||||
| -rw-r--r-- | sql/sql_base.cc | 58 | ||||
| -rw-r--r-- | sql/sql_class.cc | 4 | ||||
| -rw-r--r-- | sql/sql_class.h | 32 | ||||
| -rw-r--r-- | sql/sql_cursor.cc | 5 | ||||
| -rw-r--r-- | sql/sql_delete.cc | 38 | ||||
| -rw-r--r-- | sql/sql_do.cc | 12 | ||||
| -rw-r--r-- | sql/sql_insert.cc | 42 | ||||
| -rw-r--r-- | sql/sql_load.cc | 10 | ||||
| -rw-r--r-- | sql/sql_parse.cc | 27 | ||||
| -rw-r--r-- | sql/sql_partition.cc | 40 | ||||
| -rw-r--r-- | sql/sql_table.cc | 21 | ||||
| -rw-r--r-- | sql/sql_udf.cc | 9 | ||||
| -rw-r--r-- | sql/sql_update.cc | 29 |
21 files changed, 875 insertions, 243 deletions
diff --git a/sql/ha_ndbcluster_binlog.cc b/sql/ha_ndbcluster_binlog.cc index 07b0d907229..f5fda16b391 100644 --- a/sql/ha_ndbcluster_binlog.cc +++ b/sql/ha_ndbcluster_binlog.cc @@ -283,6 +283,7 @@ static void run_query(THD *thd, char *buf, char *end, thd_ndb->m_error_code, (int) thd->is_error(), thd->is_slave_error); } + close_thread_tables(thd); /* XXX: this code is broken. mysql_parse()/mysql_reset_thd_for_next_command() can not be called from within a statement, and diff --git a/sql/handler.cc b/sql/handler.cc index 95d63740039..d831a1c53a0 100644 --- a/sql/handler.cc +++ b/sql/handler.cc @@ -576,6 +576,295 @@ void ha_close_connection(THD* thd) ======================= TRANSACTIONS ===================================*/ /** + Transaction handling in the server + ================================== + + In each client connection, MySQL maintains two transactional + states: + - a statement transaction, + - a standard, also called normal transaction. + + Historical note + --------------- + "Statement transaction" is a non-standard term that comes + from the times when MySQL supported BerkeleyDB storage engine. + + First of all, it should be said that in BerkeleyDB auto-commit + mode auto-commits operations that are atomic to the storage + engine itself, such as a write of a record, and are too + high-granular to be atomic from the application perspective + (MySQL). One SQL statement could involve many BerkeleyDB + auto-committed operations and thus BerkeleyDB auto-commit was of + little use to MySQL. + + Secondly, instead of SQL standard savepoints, BerkeleyDB + provided the concept of "nested transactions". In a nutshell, + transactions could be arbitrarily nested, but when the parent + transaction was committed or aborted, all its child (nested) + transactions were handled committed or aborted as well. + Commit of a nested transaction, in turn, made its changes + visible, but not durable: it destroyed the nested transaction, + all its changes would become available to the parent and + currently active nested transactions of this parent. + + So the mechanism of nested transactions was employed to + provide "all or nothing" guarantee of SQL statements + required by the standard. + A nested transaction would be created at start of each SQL + statement, and destroyed (committed or aborted) at statement + end. Such nested transaction was internally referred to as + a "statement transaction" and gave birth to the term. + + <Historical note ends> + + Since then a statement transaction is started for each statement + that accesses transactional tables or uses the binary log. If + the statement succeeds, the statement transaction is committed. + If the statement fails, the transaction is rolled back. Commits + of statement transactions are not durable -- each such + transaction is nested in the normal transaction, and if the + normal transaction is rolled back, the effects of all enclosed + statement transactions are undone as well. Technically, + a statement transaction can be viewed as a savepoint which is + maintained automatically in order to make effects of one + statement atomic. + + The normal transaction is started by the user and is ended + usually upon a user request as well. The normal transaction + encloses transactions of all statements issued between + its beginning and its end. + In autocommit mode, the normal transaction is equivalent + to the statement transaction. + + Since MySQL supports PSEA (pluggable storage engine + architecture), more than one transactional engine can be + active at a time. Hence transactions, from the server + point of view, are always distributed. In particular, + transactional state is maintained independently for each + engine. In order to commit a transaction the two phase + commit protocol is employed. + + Not all statements are executed in context of a transaction. + Administrative and status information statements do not modify + engine data, and thus do not start a statement transaction and + also have no effect on the normal transaction. Examples of such + statements are SHOW STATUS and RESET SLAVE. + + Similarly DDL statements are not transactional, + and therefore a transaction is [almost] never started for a DDL + statement. The difference between a DDL statement and a purely + administrative statement though is that a DDL statement always + commits the current transaction before proceeding, if there is + any. + + At last, SQL statements that work with non-transactional + engines also have no effect on the transaction state of the + connection. Even though they are written to the binary log, + and the binary log is, overall, transactional, the writes + are done in "write-through" mode, directly to the binlog + file, followed with a OS cache sync, in other words, + bypassing the binlog undo log (translog). + They do not commit the current normal transaction. + A failure of a statement that uses non-transactional tables + would cause a rollback of the statement transaction, but + in case there no non-transactional tables are used, + no statement transaction is started. + + Data layout + ----------- + + The server stores its transaction-related data in + thd->transaction. This structure has two members of type + THD_TRANS. These members correspond to the statement and + normal transactions respectively: + + - thd->transaction.stmt contains a list of engines + that are participating in the given statement + - thd->transaction.all contains a list of engines that + have participated in any of the statement transactions started + within the context of the normal transaction. + Each element of the list contains a pointer to the storage + engine, engine-specific transactional data, and engine-specific + transaction flags. + + In autocommit mode thd->transaction.all is empty. + Instead, data of thd->transaction.stmt is + used to commit/rollback the normal transaction. + + The list of registered engines has a few important properties: + - no engine is registered in the list twice + - engines are present in the list a reverse temporal order -- + new participants are always added to the beginning of the list. + + Transaction life cycle + ---------------------- + + When a new connection is established, thd->transaction + members are initialized to an empty state. + If a statement uses any tables, all affected engines + are registered in the statement engine list. In + non-autocommit mode, the same engines are registered in + the normal transaction list. + At the end of the statement, the server issues a commit + or a roll back for all engines in the statement list. + At this point transaction flags of an engine, if any, are + propagated from the statement list to the list of the normal + transaction. + When commit/rollback is finished, the statement list is + cleared. It will be filled in again by the next statement, + and emptied again at the next statement's end. + + The normal transaction is committed in a similar way + (by going over all engines in thd->transaction.all list) + but at different times: + - upon COMMIT SQL statement is issued by the user + - implicitly, by the server, at the beginning of a DDL statement + or SET AUTOCOMMIT={0|1} statement. + + The normal transaction can be rolled back as well: + - if the user has requested so, by issuing ROLLBACK SQL + statement + - if one of the storage engines requested a rollback + by setting thd->transaction_rollback_request. This may + happen in case, e.g., when the transaction in the engine was + chosen a victim of the internal deadlock resolution algorithm + and rolled back internally. When such a situation happens, there + is little the server can do and the only option is to rollback + transactions in all other participating engines. In this case + the rollback is accompanied by an error sent to the user. + + As follows from the use cases above, the normal transaction + is never committed when there is an outstanding statement + transaction. In most cases there is no conflict, since + commits of the normal transaction are issued by a stand-alone + administrative or DDL statement, thus no outstanding statement + transaction of the previous statement exists. Besides, + all statements that manipulate with the normal transaction + are prohibited in stored functions and triggers, therefore + no conflicting situation can occur in a sub-statement either. + The remaining rare cases when the server explicitly has + to commit the statement transaction prior to committing the normal + one cover error-handling scenarios (see for example + SQLCOM_LOCK_TABLES). + + When committing a statement or a normal transaction, the server + either uses the two-phase commit protocol, or issues a commit + in each engine independently. The two-phase commit protocol + is used only if: + - all participating engines support two-phase commit (provide + handlerton::prepare PSEA API call) and + - transactions in at least two engines modify data (i.e. are + not read-only). + + Note that the two phase commit is used for + statement transactions, even though they are not durable anyway. + This is done to ensure logical consistency of data in a multiple- + engine transaction. + For example, imagine that some day MySQL supports unique + constraint checks deferred till the end of statement. In such + case a commit in one of the engines may yield ER_DUP_KEY, + and MySQL should be able to gracefully abort statement + transactions of other participants. + + After the normal transaction has been committed, + thd->transaction.all list is cleared. + + When a connection is closed, the current normal transaction, if + any, is rolled back. + + Roles and responsibilities + -------------------------- + + The server has no way to know that an engine participates in + the statement and a transaction has been started + in it unless the engine says so. Thus, in order to be + a part of a transaction, the engine must "register" itself. + This is done by invoking trans_register_ha() server call. + Normally the engine registers itself whenever handler::external_lock() + is called. trans_register_ha() can be invoked many times: if + an engine is already registered, the call does nothing. + In case autocommit is not set, the engine must register itself + twice -- both in the statement list and in the normal transaction + list. + In which list to register is a parameter of trans_register_ha(). + + Note, that although the registration interface in itself is + fairly clear, the current usage practice often leads to undesired + effects. E.g. since a call to trans_register_ha() in most engines + is embedded into implementation of handler::external_lock(), some + DDL statements start a transaction (at least from the server + point of view) even though they are not expected to. E.g. + CREATE TABLE does not start a transaction, since + handler::external_lock() is never called during CREATE TABLE. But + CREATE TABLE ... SELECT does, since handler::external_lock() is + called for the table that is being selected from. This has no + practical effects currently, but must be kept in mind + nevertheless. + + Once an engine is registered, the server will do the rest + of the work. + + During statement execution, whenever any of data-modifying + PSEA API methods is used, e.g. handler::write_row() or + handler::update_row(), the read-write flag is raised in the + statement transaction for the involved engine. + Currently All PSEA calls are "traced", and the data can not be + changed in a way other than issuing a PSEA call. Important: + unless this invariant is preserved the server will not know that + a transaction in a given engine is read-write and will not + involve the two-phase commit protocol! + + At the end of a statement, server call + ha_autocommit_or_rollback() is invoked. This call in turn + invokes handlerton::prepare() for every involved engine. + Prepare is followed by a call to handlerton::commit_one_phase() + If a one-phase commit will suffice, handlerton::prepare() is not + invoked and the server only calls handlerton::commit_one_phase(). + At statement commit, the statement-related read-write engine + flag is propagated to the corresponding flag in the normal + transaction. When the commit is complete, the list of registered + engines is cleared. + + Rollback is handled in a similar fashion. + + Additional notes on DDL and the normal transaction. + --------------------------------------------------- + + DDLs and operations with non-transactional engines + do not "register" in thd->transaction lists, and thus do not + modify the transaction state. Besides, each DDL in + MySQL is prefixed with an implicit normal transaction commit + (a call to end_active_trans()), and thus leaves nothing + to modify. + However, as it has been pointed out with CREATE TABLE .. SELECT, + some DDL statements can start a *new* transaction. + + Behaviour of the server in this case is currently badly + defined. + DDL statements use a form of "semantic" logging + to maintain atomicity: if CREATE TABLE .. SELECT failed, + the newly created table is deleted. + In addition, some DDL statements issue interim transaction + commits: e.g. ALTER TABLE issues a commit after data is copied + from the original table to the internal temporary table. Other + statements, e.g. CREATE TABLE ... SELECT do not always commit + after itself. + And finally there is a group of DDL statements such as + RENAME/DROP TABLE that doesn't start a new transaction + and doesn't commit. + + This diversity makes it hard to say what will happen if + by chance a stored function is invoked during a DDL -- + whether any modifications it makes will be committed or not + is not clear. Fortunately, SQL grammar of few DDLs allows + invocation of a stored function. + + A consistent behaviour is perhaps to always commit the normal + transaction after all DDLs, just like the statement transaction + is always committed at the end of all statements. +*/ + +/** Register a storage engine for a transaction. Every storage engine MUST call this function when it starts @@ -592,7 +881,7 @@ void ha_close_connection(THD* thd) void trans_register_ha(THD *thd, bool all, handlerton *ht_arg) { THD_TRANS *trans; - handlerton **ht; + Ha_trx_info *ha_info; DBUG_ENTER("trans_register_ha"); DBUG_PRINT("enter",("%s", all ? "all" : "stmt")); @@ -604,12 +893,13 @@ void trans_register_ha(THD *thd, bool all, handlerton *ht_arg) else trans= &thd->transaction.stmt; - for (ht=trans->ht; *ht; ht++) - if (*ht == ht_arg) - DBUG_VOID_RETURN; /* already registered, return */ + ha_info= thd->ha_data[ht_arg->slot].ha_info + static_cast<unsigned>(all); + + if (ha_info->is_started()) + DBUG_VOID_RETURN; /* already registered, return */ + + ha_info->register_ha(trans, ht_arg); - trans->ht[trans->nht++]=ht_arg; - DBUG_ASSERT(*ht == ht_arg); trans->no_2pc|=(ht_arg->prepare==0); if (thd->transaction.xid_state.xid.is_null()) thd->transaction.xid_state.xid.set(thd->query_id); @@ -626,18 +916,19 @@ int ha_prepare(THD *thd) { int error=0, all=1; THD_TRANS *trans=all ? &thd->transaction.all : &thd->transaction.stmt; - handlerton **ht=trans->ht; + Ha_trx_info *ha_info= trans->ha_list; DBUG_ENTER("ha_prepare"); #ifdef USING_TRANSACTIONS - if (trans->nht) + if (ha_info) { - for (; *ht; ht++) + for (; ha_info; ha_info= ha_info->next()) { int err; + handlerton *ht= ha_info->ht(); status_var_increment(thd->status_var.ha_prepare_count); - if ((*ht)->prepare) + if (ht->prepare) { - if ((err= (*(*ht)->prepare)(*ht, thd, all))) + if ((err= ht->prepare(ht, thd, all))) { my_error(ER_ERROR_DURING_COMMIT, MYF(0), err); ha_rollback_trans(thd, all); @@ -649,7 +940,7 @@ int ha_prepare(THD *thd) { push_warning_printf(thd, MYSQL_ERROR::WARN_LEVEL_WARN, ER_ILLEGAL_HA, ER(ER_ILLEGAL_HA), - ha_resolve_storage_engine_name(*ht)); + ha_resolve_storage_engine_name(ht)); } } } @@ -658,6 +949,62 @@ int ha_prepare(THD *thd) } /** + Check if we can skip the two-phase commit. + + A helper function to evaluate if two-phase commit is mandatory. + As a side effect, propagates the read-only/read-write flags + of the statement transaction to its enclosing normal transaction. + + @retval TRUE we must run a two-phase commit. Returned + if we have at least two engines with read-write changes. + @retval FALSE Don't need two-phase commit. Even if we have two + transactional engines, we can run two independent + commits if changes in one of the engines are read-only. +*/ + +static +bool +ha_check_and_coalesce_trx_read_only(THD *thd, Ha_trx_info *ha_list, + bool all) +{ + /* The number of storage engines that have actual changes. */ + unsigned rw_ha_count= 0; + Ha_trx_info *ha_info; + + for (ha_info= ha_list; ha_info; ha_info= ha_info->next()) + { + if (ha_info->is_trx_read_write()) + ++rw_ha_count; + + if (! all) + { + Ha_trx_info *ha_info_all= &thd->ha_data[ha_info->ht()->slot].ha_info[1]; + DBUG_ASSERT(ha_info != ha_info_all); + /* + Merge read-only/read-write information about statement + transaction to its enclosing normal transaction. Do this + only if in a real transaction -- that is, if we know + that ha_info_all is registered in thd->transaction.all. + Since otherwise we only clutter the normal transaction flags. + */ + if (ha_info_all->is_started()) /* FALSE if autocommit. */ + ha_info_all->coalesce_trx_with(ha_info); + } + else if (rw_ha_count > 1) + { + /* + It is a normal transaction, so we don't need to merge read/write + information up, and the need for two-phase commit has been + already established. Break the loop prematurely. + */ + break; + } + } + return rw_ha_count > 1; +} + + +/** @retval 0 ok @retval @@ -674,12 +1021,25 @@ int ha_prepare(THD *thd) int ha_commit_trans(THD *thd, bool all) { int error= 0, cookie= 0; + /* + 'all' means that this is either an explicit commit issued by + user, or an implicit commit issued by a DDL. + */ THD_TRANS *trans= all ? &thd->transaction.all : &thd->transaction.stmt; - bool is_real_trans= all || thd->transaction.all.nht == 0; - handlerton **ht= trans->ht; + bool is_real_trans= all || thd->transaction.all.ha_list == 0; + Ha_trx_info *ha_info= trans->ha_list; my_xid xid= thd->transaction.xid_state.xid.get_my_xid(); DBUG_ENTER("ha_commit_trans"); + /* + We must not commit the normal transaction if a statement + transaction is pending. Otherwise statement transaction + flags will not get propagated to its normal transaction's + counterpart. + */ + DBUG_ASSERT(thd->transaction.stmt.ha_list == NULL || + trans == &thd->transaction.stmt); + if (thd->in_sub_stmt) { /* @@ -701,8 +1061,10 @@ int ha_commit_trans(THD *thd, bool all) DBUG_RETURN(2); } #ifdef USING_TRANSACTIONS - if (trans->nht) + if (ha_info) { + bool must_2pc; + if (is_real_trans && wait_if_global_read_lock(thd, 0, 0)) { ha_rollback_trans(thd, all); @@ -727,12 +1089,26 @@ int ha_commit_trans(THD *thd, bool all) if (is_real_trans) /* not a statement commit */ thd->stmt_map.close_transient_cursors(); - if (!trans->no_2pc && trans->nht > 1) + must_2pc= ha_check_and_coalesce_trx_read_only(thd, ha_info, all); + + if (!trans->no_2pc && must_2pc) { - for (; *ht && !error; ht++) + for (; ha_info && !error; ha_info= ha_info->next()) { int err; - if ((err= (*(*ht)->prepare)(*ht, thd, all))) + handlerton *ht= ha_info->ht(); + /* + Do not call two-phase commit if this particular + transaction is read-only. This allows for simpler + implementation in engines that are always read-only. + */ + if (! ha_info->is_trx_read_write()) + continue; + /* + Sic: we know that prepare() is not NULL since otherwise + trans->no_2pc would have been set. + */ + if ((err= ht->prepare(ht, thd, all))) { my_error(ER_ERROR_DURING_COMMIT, MYF(0), err); error= 1; @@ -770,24 +1146,26 @@ int ha_commit_one_phase(THD *thd, bool all) { int error=0; THD_TRANS *trans=all ? &thd->transaction.all : &thd->transaction.stmt; - bool is_real_trans=all || thd->transaction.all.nht == 0; - handlerton **ht=trans->ht; + bool is_real_trans=all || thd->transaction.all.ha_list == 0; + Ha_trx_info *ha_info= trans->ha_list, *ha_info_next; DBUG_ENTER("ha_commit_one_phase"); #ifdef USING_TRANSACTIONS - if (trans->nht) + if (ha_info) { - for (ht=trans->ht; *ht; ht++) + for (; ha_info; ha_info= ha_info_next) { int err; - if ((err= (*(*ht)->commit)(*ht, thd, all))) + handlerton *ht= ha_info->ht(); + if ((err= ht->commit(ht, thd, all))) { my_error(ER_ERROR_DURING_COMMIT, MYF(0), err); error=1; } status_var_increment(thd->status_var.ha_commit_count); - *ht= 0; + ha_info_next= ha_info->next(); + ha_info->reset(); /* keep it conveniently zero-filled */ } - trans->nht=0; + trans->ha_list= 0; trans->no_2pc=0; if (is_real_trans) thd->transaction.xid_state.xid.null(); @@ -810,8 +1188,17 @@ int ha_rollback_trans(THD *thd, bool all) { int error=0; THD_TRANS *trans=all ? &thd->transaction.all : &thd->transaction.stmt; - bool is_real_trans=all || thd->transaction.all.nht == 0; + Ha_trx_info *ha_info= trans->ha_list, *ha_info_next; + bool is_real_trans=all || thd->transaction.all.ha_list == 0; DBUG_ENTER("ha_rollback_trans"); + + /* + We must not rollback the normal transaction if a statement + transaction is pending. + */ + DBUG_ASSERT(thd->transaction.stmt.ha_list == NULL || + trans == &thd->transaction.stmt); + if (thd->in_sub_stmt) { /* @@ -826,24 +1213,26 @@ int ha_rollback_trans(THD *thd, bool all) DBUG_RETURN(1); } #ifdef USING_TRANSACTIONS - if (trans->nht) + if (ha_info) { /* Close all cursors that can not survive ROLLBACK */ if (is_real_trans) /* not a statement commit */ thd->stmt_map.close_transient_cursors(); - for (handlerton **ht=trans->ht; *ht; ht++) + for (; ha_info; ha_info= ha_info_next) { int err; - if ((err= (*(*ht)->rollback)(*ht, thd, all))) + handlerton *ht= ha_info->ht(); + if ((err= ht->rollback(ht, thd, all))) { // cannot happen my_error(ER_ERROR_DURING_ROLLBACK, MYF(0), err); error=1; } status_var_increment(thd->status_var.ha_rollback_count); - *ht= 0; + ha_info_next= ha_info->next(); + ha_info->reset(); /* keep it conveniently zero-filled */ } - trans->nht=0; + trans->ha_list= 0; trans->no_2pc=0; if (is_real_trans) thd->transaction.xid_state.xid.null(); @@ -889,17 +1278,19 @@ int ha_autocommit_or_rollback(THD *thd, int error) { DBUG_ENTER("ha_autocommit_or_rollback"); #ifdef USING_TRANSACTIONS - if (thd->transaction.stmt.nht) + if (thd->transaction.stmt.ha_list) { if (!error) { - if (ha_commit_stmt(thd)) + if (ha_commit_trans(thd, 0)) error=1; } - else if (thd->transaction_rollback_request && !thd->in_sub_stmt) - (void) ha_rollback(thd); - else - (void) ha_rollback_stmt(thd); + else + { + (void) ha_rollback_trans(thd, 0); + if (thd->transaction_rollback_request && !thd->in_sub_stmt) + (void) ha_rollback(thd); + } thd->variables.tx_isolation=thd->session_tx_isolation; } @@ -1246,43 +1637,49 @@ int ha_rollback_to_savepoint(THD *thd, SAVEPOINT *sv) int error=0; THD_TRANS *trans= (thd->in_sub_stmt ? &thd->transaction.stmt : &thd->transaction.all); - handlerton **ht=trans->ht, **end_ht; + Ha_trx_info *ha_info, *ha_info_next; + DBUG_ENTER("ha_rollback_to_savepoint"); - trans->nht=sv->nht; trans->no_2pc=0; - end_ht=ht+sv->nht; /* rolling back to savepoint in all storage engines that were part of the transaction when the savepoint was set */ - for (; ht < end_ht; ht++) + for (ha_info= sv->ha_list; ha_info; ha_info= ha_info->next()) { int err; - DBUG_ASSERT((*ht)->savepoint_set != 0); - if ((err= (*(*ht)->savepoint_rollback)(*ht, thd, (uchar *)(sv+1)+(*ht)->savepoint_offset))) + handlerton *ht= ha_info->ht(); + DBUG_ASSERT(ht); + DBUG_ASSERT(ht->savepoint_set != 0); + if ((err= ht->savepoint_rollback(ht, thd, + (uchar *)(sv+1)+ht->savepoint_offset))) { // cannot happen my_error(ER_ERROR_DURING_ROLLBACK, MYF(0), err); error=1; } status_var_increment(thd->status_var.ha_savepoint_rollback_count); - trans->no_2pc|=(*ht)->prepare == 0; + trans->no_2pc|= ht->prepare == 0; } /* rolling back the transaction in all storage engines that were not part of the transaction when the savepoint was set */ - for (; *ht ; ht++) + for (ha_info= trans->ha_list; ha_info != sv->ha_list; + ha_info= ha_info_next) { int err; - if ((err= (*(*ht)->rollback)(*ht, thd, !thd->in_sub_stmt))) + handlerton *ht= ha_info->ht(); + if ((err= ht->rollback(ht, thd, !thd->in_sub_stmt))) { // cannot happen my_error(ER_ERROR_DURING_ROLLBACK, MYF(0), err); error=1; } status_var_increment(thd->status_var.ha_rollback_count); - *ht=0; // keep it conveniently zero-filled + ha_info_next= ha_info->next(); + ha_info->reset(); /* keep it conveniently zero-filled */ } + trans->ha_list= sv->ha_list; DBUG_RETURN(error); } @@ -1297,26 +1694,32 @@ int ha_savepoint(THD *thd, SAVEPOINT *sv) int error=0; THD_TRANS *trans= (thd->in_sub_stmt ? &thd->transaction.stmt : &thd->transaction.all); - handlerton **ht=trans->ht; + Ha_trx_info *ha_info= trans->ha_list; DBUG_ENTER("ha_savepoint"); #ifdef USING_TRANSACTIONS - for (; *ht; ht++) + for (; ha_info; ha_info= ha_info->next()) { int err; - if (! (*ht)->savepoint_set) + handlerton *ht= ha_info->ht(); + DBUG_ASSERT(ht); + if (! ht->savepoint_set) { my_error(ER_CHECK_NOT_IMPLEMENTED, MYF(0), "SAVEPOINT"); error=1; break; } - if ((err= (*(*ht)->savepoint_set)(*ht, thd, (uchar *)(sv+1)+(*ht)->savepoint_offset))) + if ((err= ht->savepoint_set(ht, thd, (uchar *)(sv+1)+ht->savepoint_offset))) { // cannot happen my_error(ER_GET_ERRNO, MYF(0), err); error=1; } status_var_increment(thd->status_var.ha_savepoint_count); } - sv->nht=trans->nht; + /* + Remember the list of registered storage engines. All new + engines are prepended to the beginning of the list. + */ + sv->ha_list= trans->ha_list; #endif /* USING_TRANSACTIONS */ DBUG_RETURN(error); } @@ -1324,20 +1727,19 @@ int ha_savepoint(THD *thd, SAVEPOINT *sv) int ha_release_savepoint(THD *thd, SAVEPOINT *sv) { int error=0; - THD_TRANS *trans= (thd->in_sub_stmt ? &thd->transaction.stmt : - &thd->transaction.all); - handlerton **ht=trans->ht, **end_ht; + Ha_trx_info *ha_info= sv->ha_list; DBUG_ENTER("ha_release_savepoint"); - end_ht=ht+sv->nht; - for (; ht < end_ht; ht++) + for (; ha_info; ha_info= ha_info->next()) { int err; - if (!(*ht)->savepoint_release) + handlerton *ht= ha_info->ht(); + /* Savepoint life time is enclosed into transaction life time. */ + DBUG_ASSERT(ht); + if (!ht->savepoint_release) continue; - if ((err= (*(*ht)->savepoint_release)(*ht, thd, - (uchar *)(sv+1)+ - (*ht)->savepoint_offset))) + if ((err= ht->savepoint_release(ht, thd, + (uchar *)(sv+1) + ht->savepoint_offset))) { // cannot happen my_error(ER_GET_ERRNO, MYF(0), err); error=1; @@ -2506,6 +2908,36 @@ int handler::ha_check(THD *thd, HA_CHECK_OPT *check_opt) return update_frm_version(table); } +/** + A helper function to mark a transaction read-write, + if it is started. +*/ + +inline +void +handler::mark_trx_read_write() +{ + Ha_trx_info *ha_info= &ha_thd()->ha_data[ht->slot].ha_info[0]; + /* + When a storage engine method is called, the transaction must + have been started, unless it's a DDL call, for which the + storage engine starts the transaction internally, and commits + it internally, without registering in the ha_list. + Unfortunately here we can't know know for sure if the engine + has registered the transaction or not, so we must check. + */ + if (ha_info->is_started()) + { + DBUG_ASSERT(has_transactions()); + /* + table_share can be NULL in ha_delete_table(). See implementation + of standalone function ha_delete_table() in sql_base.cc. + */ + if (table_share == NULL || table_share->tmp_table == NO_TMP_TABLE) + ha_info->set_trx_read_write(); + } +} + /** Repair table: public interface. @@ -2516,6 +2948,9 @@ int handler::ha_check(THD *thd, HA_CHECK_OPT *check_opt) int handler::ha_repair(THD* thd, HA_CHECK_OPT* check_opt) { int result; + + mark_trx_read_write(); + if ((result= repair(thd, check_opt))) return result; return update_frm_version(table); @@ -2532,6 +2967,8 @@ int handler::ha_bulk_update_row(const uchar *old_data, uchar *new_data, uint *dup_key_found) { + mark_trx_read_write(); + return bulk_update_row(old_data, new_data, dup_key_found); } @@ -2545,6 +2982,8 @@ handler::ha_bulk_update_row(const uchar *old_data, uchar *new_data, int handler::ha_delete_all_rows() { + mark_trx_read_write(); + return delete_all_rows(); } @@ -2558,6 +2997,8 @@ handler::ha_delete_all_rows() int handler::ha_reset_auto_increment(ulonglong value) { + mark_trx_read_write(); + return reset_auto_increment(value); } @@ -2571,6 +3012,8 @@ handler::ha_reset_auto_increment(ulonglong value) int handler::ha_backup(THD* thd, HA_CHECK_OPT* check_opt) { + mark_trx_read_write(); + return backup(thd, check_opt); } @@ -2584,6 +3027,8 @@ handler::ha_backup(THD* thd, HA_CHECK_OPT* check_opt) int handler::ha_restore(THD* thd, HA_CHECK_OPT* check_opt) { + mark_trx_read_write(); + return restore(thd, check_opt); } @@ -2597,6 +3042,8 @@ handler::ha_restore(THD* thd, HA_CHECK_OPT* check_opt) int handler::ha_optimize(THD* thd, HA_CHECK_OPT* check_opt) { + mark_trx_read_write(); + return optimize(thd, check_opt); } @@ -2610,6 +3057,8 @@ handler::ha_optimize(THD* thd, HA_CHECK_OPT* check_opt) int handler::ha_analyze(THD* thd, HA_CHECK_OPT* check_opt) { + mark_trx_read_write(); + return analyze(thd, check_opt); } @@ -2623,6 +3072,8 @@ handler::ha_analyze(THD* thd, HA_CHECK_OPT* check_opt) bool handler::ha_check_and_repair(THD *thd) { + mark_trx_read_write(); + return check_and_repair(thd); } @@ -2636,6 +3087,8 @@ handler::ha_check_and_repair(THD *thd) int handler::ha_disable_indexes(uint mode) { + mark_trx_read_write(); + return disable_indexes(mode); } @@ -2649,6 +3102,8 @@ handler::ha_disable_indexes(uint mode) int handler::ha_enable_indexes(uint mode) { + mark_trx_read_write(); + return enable_indexes(mode); } @@ -2662,6 +3117,8 @@ handler::ha_enable_indexes(uint mode) int handler::ha_discard_or_import_tablespace(my_bool discard) { + mark_trx_read_write(); + return discard_or_import_tablespace(discard); } @@ -2677,6 +3134,8 @@ handler::ha_discard_or_import_tablespace(my_bool discard) void handler::ha_prepare_for_alter() { + mark_trx_read_write(); + prepare_for_alter(); } @@ -2690,6 +3149,8 @@ handler::ha_prepare_for_alter() int handler::ha_rename_table(const char *from, const char *to) { + mark_trx_read_write(); + return rename_table(from, to); } @@ -2703,6 +3164,8 @@ handler::ha_rename_table(const char *from, const char *to) int handler::ha_delete_table(const char *name) { + mark_trx_read_write(); + return delete_table(name); } @@ -2716,6 +3179,8 @@ handler::ha_delete_table(const char *name) void handler::ha_drop_table(const char *name) { + mark_trx_read_write(); + return drop_table(name); } @@ -2729,6 +3194,8 @@ handler::ha_drop_table(const char *name) int handler::ha_create(const char *name, TABLE *form, HA_CREATE_INFO *info) { + mark_trx_read_write(); + return create(name, form, info); } @@ -2743,6 +3210,8 @@ int handler::ha_create_handler_files(const char *name, const char *old_name, int action_flag, HA_CREATE_INFO *info) { + mark_trx_read_write(); + return create_handler_files(name, old_name, action_flag, info); } @@ -2761,6 +3230,8 @@ handler::ha_change_partitions(HA_CREATE_INFO *create_info, const uchar *pack_frm_data, size_t pack_frm_len) { + mark_trx_read_write(); + return change_partitions(create_info, path, copied, deleted, pack_frm_data, pack_frm_len); } @@ -2775,6 +3246,8 @@ handler::ha_change_partitions(HA_CREATE_INFO *create_info, int handler::ha_drop_partitions(const char *path) { + mark_trx_read_write(); + return drop_partitions(path); } @@ -2788,6 +3261,8 @@ handler::ha_drop_partitions(const char *path) int handler::ha_rename_partitions(const char *path) { + mark_trx_read_write(); + return rename_partitions(path); } @@ -2801,6 +3276,8 @@ handler::ha_rename_partitions(const char *path) int handler::ha_optimize_partitions(THD *thd) { + mark_trx_read_write(); + return optimize_partitions(thd); } @@ -2814,6 +3291,8 @@ handler::ha_optimize_partitions(THD *thd) int handler::ha_analyze_partitions(THD *thd) { + mark_trx_read_write(); + return analyze_partitions(thd); } @@ -2827,6 +3306,8 @@ handler::ha_analyze_partitions(THD *thd) int handler::ha_check_partitions(THD *thd) { + mark_trx_read_write(); + return check_partitions(thd); } @@ -2840,6 +3321,8 @@ handler::ha_check_partitions(THD *thd) int handler::ha_repair_partitions(THD *thd) { + mark_trx_read_write(); + return repair_partitions(thd); } @@ -2866,7 +3349,7 @@ int ha_enable_transaction(THD *thd, bool on) is an optimization hint that storage engine is free to ignore. So, let's commit an open transaction (if any) now. */ - if (!(error= ha_commit_stmt(thd))) + if (!(error= ha_commit_trans(thd, 0))) error= end_trans(thd, COMMIT); } DBUG_RETURN(error); @@ -4042,6 +4525,9 @@ int handler::ha_write_row(uchar *buf) { int error; DBUG_ENTER("handler::ha_write_row"); + + mark_trx_read_write(); + if (unlikely(error= write_row(buf))) DBUG_RETURN(error); if (unlikely(error= binlog_log_row<Write_rows_log_event>(table, 0, buf))) @@ -4060,6 +4546,8 @@ int handler::ha_update_row(const uchar *old_data, uchar *new_data) */ DBUG_ASSERT(new_data == table->record[0]); + mark_trx_read_write(); + if (unlikely(error= update_row(old_data, new_data))) return error; if (unlikely(error= binlog_log_row<Update_rows_log_event>(table, old_data, new_data))) @@ -4070,6 +4558,9 @@ int handler::ha_update_row(const uchar *old_data, uchar *new_data) int handler::ha_delete_row(const uchar *buf) { int error; + + mark_trx_read_write(); + if (unlikely(error= delete_row(buf))) return error; if (unlikely(error= binlog_log_row<Delete_rows_log_event>(table, buf, 0))) diff --git a/sql/handler.h b/sql/handler.h index b3a4b408589..9800f4974c3 100644 --- a/sql/handler.h +++ b/sql/handler.h @@ -721,14 +721,14 @@ struct handlerton #define HTON_SUPPORT_LOG_TABLES (1 << 7) //Engine supports log tables #define HTON_NO_PARTITION (1 << 8) //You can not partition these tables -typedef struct st_thd_trans +class Ha_trx_info; + +struct THD_TRANS { - /* number of entries in the ht[] */ - uint nht; /* true is not all entries in the ht[] support 2pc */ bool no_2pc; - /* storage engines that registered themselves for this transaction */ - handlerton *ht[MAX_HA]; + /* storage engines that registered in this transaction */ + Ha_trx_info *ha_list; /* The purpose of this flag is to keep track of non-transactional tables that were modified in scope of: @@ -758,7 +758,106 @@ typedef struct st_thd_trans saved value. */ bool modified_non_trans_table; -} THD_TRANS; + + void reset() { no_2pc= FALSE; modified_non_trans_table= FALSE; } +}; + + +/** + Either statement transaction or normal transaction - related + thread-specific storage engine data. + + If a storage engine participates in a statement/transaction, + an instance of this class is present in + thd->transaction.{stmt|all}.ha_list. The addition to + {stmt|all}.ha_list is made by trans_register_ha(). + + When it's time to commit or rollback, each element of ha_list + is used to access storage engine's prepare()/commit()/rollback() + methods, and also to evaluate if a full two phase commit is + necessary. + + @sa General description of transaction handling in handler.cc. +*/ + +class Ha_trx_info +{ +public: + /** Register this storage engine in the given transaction context. */ + void register_ha(THD_TRANS *trans, handlerton *ht_arg) + { + DBUG_ASSERT(m_flags == 0); + DBUG_ASSERT(m_ht == NULL); + DBUG_ASSERT(m_next == NULL); + + m_ht= ht_arg; + m_flags= (int) TRX_READ_ONLY; /* Assume read-only at start. */ + + m_next= trans->ha_list; + trans->ha_list= this; + } + + /** Clear, prepare for reuse. */ + void reset() + { + m_next= NULL; + m_ht= NULL; + m_flags= 0; + } + + Ha_trx_info() { reset(); } + + void set_trx_read_write() + { + DBUG_ASSERT(is_started()); + m_flags|= (int) TRX_READ_WRITE; + } + bool is_trx_read_write() const + { + DBUG_ASSERT(is_started()); + return m_flags & (int) TRX_READ_WRITE; + } + bool is_started() const { return m_ht != NULL; } + /** Mark this transaction read-write if the argument is read-write. */ + void coalesce_trx_with(const Ha_trx_info *stmt_trx) + { + /* + Must be called only after the transaction has been started. + Can be called many times, e.g. when we have many + read-write statements in a transaction. + */ + DBUG_ASSERT(is_started()); + if (stmt_trx->is_trx_read_write()) + set_trx_read_write(); + } + Ha_trx_info *next() const + { + DBUG_ASSERT(is_started()); + return m_next; + } + handlerton *ht() const + { + DBUG_ASSERT(is_started()); + return m_ht; + } +private: + enum { TRX_READ_ONLY= 0, TRX_READ_WRITE= 1 }; + /** Auxiliary, used for ha_list management */ + Ha_trx_info *m_next; + /** + Although a given Ha_trx_info instance is currently always used + for the same storage engine, 'ht' is not-NULL only when the + corresponding storage is a part of a transaction. + */ + handlerton *m_ht; + /** + Transaction flags related to this engine. + Not-null only if this instance is a part of transaction. + May assume a combination of enum values above. + */ + uchar m_flags; +}; + enum enum_tx_isolation { ISO_READ_UNCOMMITTED, ISO_READ_COMMITTED, ISO_REPEATABLE_READ, ISO_SERIALIZABLE}; @@ -1640,8 +1739,15 @@ protected: provide useful functionality. */ virtual int rename_table(const char *from, const char *to); + /** + Delete a table in the engine. Called for base as well as temporary + tables. + */ virtual int delete_table(const char *name); private: + /* Private helpers */ + inline void mark_trx_read_write(); +private: /* Low-level primitives for storage engines. These should be overridden by the storage engine class. To call these methods, use @@ -1821,9 +1927,7 @@ extern TYPELIB tx_isolation_typelib; extern TYPELIB myisam_stats_method_typelib; extern ulong total_ha, total_ha_2pc; - /* Wrapper functions */ -#define ha_commit_stmt(thd) (ha_commit_trans((thd), FALSE)) -#define ha_rollback_stmt(thd) (ha_rollback_trans((thd), FALSE)) + /* Wrapper functions */ #define ha_commit(thd) (ha_commit_trans((thd), TRUE)) #define ha_rollback(thd) (ha_rollback_trans((thd), TRUE)) diff --git a/sql/log.cc b/sql/log.cc index 3a09acd8fca..af9359ed6cc 100644 --- a/sql/log.cc +++ b/sql/log.cc @@ -3332,6 +3332,16 @@ THD::binlog_start_trans_and_stmt() if (options & (OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN)) trans_register_ha(this, TRUE, binlog_hton); trans_register_ha(this, FALSE, binlog_hton); + /* + Mark statement transaction as read/write. We never start + a binary log transaction and keep it read-only, + therefore it's best to mark the transaction read/write just + at the same time we start it. + Not necessary to mark the normal transaction read/write + since the statement-level flag will be propagated automatically + inside ha_commit_trans. + */ + ha_data[binlog_hton->slot].ha_info[0].set_trx_read_write(); } DBUG_VOID_RETURN; } diff --git a/sql/log_event.cc b/sql/log_event.cc index cf03dd5bf44..0d5aefd4422 100644 --- a/sql/log_event.cc +++ b/sql/log_event.cc @@ -2898,7 +2898,7 @@ int Format_description_log_event::do_apply_event(Relay_log_info const *rli) original place when it comes to us; we'll know this by checking log_pos ("artificial" events have log_pos == 0). */ - if (!artificial_event && created && thd->transaction.all.nht) + if (!artificial_event && created && thd->transaction.all.ha_list) { /* This is not an error (XA is safe), just an information */ rli->report(INFORMATION_LEVEL, 0, diff --git a/sql/rpl_injector.cc b/sql/rpl_injector.cc index aa3020c42be..684655d1c3b 100644 --- a/sql/rpl_injector.cc +++ b/sql/rpl_injector.cc @@ -62,6 +62,26 @@ int injector::transaction::commit() { DBUG_ENTER("injector::transaction::commit()"); m_thd->binlog_flush_pending_rows_event(true); + /* + Cluster replication does not preserve statement or + transaction boundaries of the master. Instead, a new + transaction on replication slave is started when a new GCI + (global checkpoint identifier) is issued, and is committed + when the last event of the check point has been received and + processed. This ensures consistency of each cluster in + cluster replication, and there is no requirement for stronger + consistency: MySQL replication is asynchronous with other + engines as well. + + A practical consequence of that is that row level replication + stream passed through the injector thread never contains + COMMIT events. + Here we should preserve the server invariant that there is no + outstanding statement transaction when the normal transaction + is committed by committing the statement transaction + explicitly. + */ + ha_autocommit_or_rollback(m_thd, 0); end_trans(m_thd, COMMIT); DBUG_RETURN(0); } diff --git a/sql/sp.cc b/sql/sp.cc index 99ffc18deea..ae738cd0a71 100644 --- a/sql/sp.cc +++ b/sql/sp.cc @@ -665,8 +665,16 @@ sp_returns_type(THD *thd, String &result, sp_head *sp) (TYPE_ENUM_PROCEDURE or TYPE_ENUM_FUNCTION). @param sp Stored routine object to store. - @return Error code. SP_OK is returned on success. Other SP_ constants are - used to indicate about errors. + @note Opens and closes the thread tables. Therefore assumes + that there are no locked tables in this thread at the time of + invocation. + Unlike some other DDL statements, *does* close the tables + in the end, since the call to this function is normally + followed by an implicit grant (sp_grant_privileges()) + and this subsequent call opens and closes mysql.procs_priv. + + @return Error code. SP_OK is returned on success. Other + SP_ constants are used to indicate about errors. */ int @@ -1223,7 +1231,13 @@ done: } -/* Drop all routines in database 'db' */ +/** + Drop all routines in database 'db' + + @note Close the thread tables, the calling code might want to + delete from other system tables afterwards. +*/ + int sp_drop_db_routines(THD *thd, char *db) { diff --git a/sql/sp_head.cc b/sql/sp_head.cc index 8e6f5a96640..1d443b70737 100644 --- a/sql/sp_head.cc +++ b/sql/sp_head.cc @@ -2700,6 +2700,7 @@ sp_lex_keeper::reset_lex_and_exec_core(THD *thd, uint *nextp, m_lex->unit.cleanup(); thd_proc_info(thd, "closing tables"); + /* Here we also commit or rollback the current statement. */ close_thread_tables(thd); thd_proc_info(thd, 0); diff --git a/sql/sql_base.cc b/sql/sql_base.cc index e5a5b51fcf6..4a417dc05b1 100644 --- a/sql/sql_base.cc +++ b/sql/sql_base.cc @@ -1324,29 +1324,45 @@ void close_thread_tables(THD *thd) Mark all temporary tables used by this statement as free for reuse. */ mark_temp_tables_as_free_for_reuse(thd); - - if (thd->locked_tables || prelocked_mode) + /* + Let us commit transaction for statement. Since in 5.0 we only have + one statement transaction and don't allow several nested statement + transactions this call will do nothing if we are inside of stored + function or trigger (i.e. statement transaction is already active and + does not belong to statement for which we do close_thread_tables()). + TODO: This should be fixed in later releases. + */ + if (!(thd->state_flags & Open_tables_state::BACKUPS_AVAIL)) { + thd->main_da.can_overwrite_status= TRUE; + ha_autocommit_or_rollback(thd, thd->is_error()); + thd->main_da.can_overwrite_status= FALSE; + /* - Let us commit transaction for statement. Since in 5.0 we only have - one statement transaction and don't allow several nested statement - transactions this call will do nothing if we are inside of stored - function or trigger (i.e. statement transaction is already active and - does not belong to statement for which we do close_thread_tables()). - TODO: This should be fixed in later releases. + Reset transaction state, but only if we're not inside a + sub-statement of a prelocked statement. */ - ha_commit_stmt(thd); + if (! prelocked_mode || thd->lex->requires_prelocking()) + thd->transaction.stmt.reset(); + } + + if (thd->locked_tables || prelocked_mode) + { /* Ensure we are calling ha_reset() for all used tables */ mark_used_tables_as_free_for_reuse(thd, thd->open_tables); - /* We are under simple LOCK TABLES so should not do anything else. */ + /* + We are under simple LOCK TABLES or we're inside a sub-statement + of a prelocked statement, so should not do anything else. + */ if (!prelocked_mode || !thd->lex->requires_prelocking()) DBUG_VOID_RETURN; /* - We are in prelocked mode, so we have to leave it now with doing - implicit UNLOCK TABLES if need. + We are in the top-level statement of a prelocked statement, + so we have to leave the prelocked mode now with doing implicit + UNLOCK TABLES if needed. */ DBUG_PRINT("info",("thd->prelocked_mode= NON_PRELOCKED")); thd->prelocked_mode= NON_PRELOCKED; @@ -1375,19 +1391,6 @@ void close_thread_tables(THD *thd) thd->lock=0; } /* - assume handlers auto-commit (if some doesn't - transaction handling - in MySQL should be redesigned to support it; it's a big change, - and it's not worth it - better to commit explicitly only writing - transactions, read-only ones should better take care of themselves. - saves some work in 2pc too) - see also sql_parse.cc - dispatch_command() - */ - if (!(thd->state_flags & Open_tables_state::BACKUPS_AVAIL)) - bzero(&thd->transaction.stmt, sizeof(thd->transaction.stmt)); - if (!thd->active_transaction()) - thd->transaction.xid_state.xid.null(); - - /* Note that we need to hold LOCK_open while changing the open_tables list. Another thread may work on it. (See: remove_table_from_cache(), mysql_wait_completed_table()) @@ -5059,10 +5062,7 @@ int decide_logging_format(THD *thd, TABLE_LIST *tables) DBUG_PRINT("info", ("error: %d", error)); if (error) - { - ha_rollback_stmt(thd); return -1; - } /* We switch to row-based format if we are in mixed mode and one of @@ -5216,7 +5216,6 @@ int lock_tables(THD *thd, TABLE_LIST *tables, uint count, bool *need_reopen) table->table->query_id= thd->query_id; if (check_lock_and_start_stmt(thd, table->table, table->lock_type)) { - ha_rollback_stmt(thd); mysql_unlock_tables(thd, thd->locked_tables); thd->locked_tables= 0; thd->options&= ~(OPTION_TABLE_LOCK); @@ -5251,7 +5250,6 @@ int lock_tables(THD *thd, TABLE_LIST *tables, uint count, bool *need_reopen) if (!table->placeholder() && check_lock_and_start_stmt(thd, table->table, table->lock_type)) { - ha_rollback_stmt(thd); DBUG_RETURN(-1); } } diff --git a/sql/sql_class.cc b/sql/sql_class.cc index 75376c53f68..2781eec4eaf 100644 --- a/sql/sql_class.cc +++ b/sql/sql_class.cc @@ -264,7 +264,7 @@ const char *set_thd_proc_info(THD *thd, const char *info, extern "C" void **thd_ha_data(const THD *thd, const struct handlerton *hton) { - return (void **) thd->ha_data + hton->slot; + return (void **) &thd->ha_data[hton->slot].ha_ptr; } extern "C" @@ -2513,7 +2513,7 @@ bool select_dumpvar::send_data(List<Item> &items) suv->update(); } } - DBUG_RETURN(0); + DBUG_RETURN(thd->is_error()); } bool select_dumpvar::send_eof() diff --git a/sql/sql_class.h b/sql/sql_class.h index e8f28b19213..2a9941f78d0 100644 --- a/sql/sql_class.h +++ b/sql/sql_class.h @@ -686,7 +686,8 @@ private: struct st_savepoint { struct st_savepoint *prev; char *name; - uint length, nht; + uint length; + Ha_trx_info *ha_list; }; enum xa_states {XA_NOTR=0, XA_ACTIVE, XA_IDLE, XA_PREPARED}; @@ -1093,6 +1094,33 @@ private: /** + Storage engine specific thread local data. +*/ + +struct Ha_data +{ + /** + Storage engine specific thread local data. + Lifetime: one user connection. + */ + void *ha_ptr; + /** + 0: Life time: one statement within a transaction. If @@autocommit is + on, also represents the entire transaction. + @sa trans_register_ha() + + 1: Life time: one transaction within a connection. + If the storage engine does not participate in a transaction, + this should not be used. + @sa trans_register_ha() + */ + Ha_trx_info ha_info[2]; + + Ha_data() :ha_ptr(NULL) {} +}; + + +/** @class THD For each client connection we create a separate thread with THD serving as a thread/connection descriptor @@ -1231,7 +1259,7 @@ public: uint in_sub_stmt; /* container for handler's private per-connection data */ - void *ha_data[MAX_HA]; + Ha_data ha_data[MAX_HA]; #ifndef MYSQL_CLIENT int binlog_setup_trx_data(); diff --git a/sql/sql_cursor.cc b/sql/sql_cursor.cc index 2301b561797..6f2bac73dc4 100644 --- a/sql/sql_cursor.cc +++ b/sql/sql_cursor.cc @@ -322,9 +322,10 @@ Sensitive_cursor::post_open(THD *thd) close_at_commit= FALSE; /* reset in case we're reusing the cursor */ info= &ht_info[0]; - for (handlerton **pht= thd->transaction.stmt.ht; *pht; pht++) + for (Ha_trx_info *ha_trx_info= thd->transaction.stmt.ha_list; + ha_trx_info; ha_trx_info= ha_trx_info->next()) { - handlerton *ht= *pht; + handlerton *ht= ha_trx_info->ht(); close_at_commit|= test(ht->flags & HTON_CLOSE_CURSORS_AT_COMMIT); if (ht->create_cursor_read_view) { diff --git a/sql/sql_delete.cc b/sql/sql_delete.cc index c6435387f44..234ee35ecca 100644 --- a/sql/sql_delete.cc +++ b/sql/sql_delete.cc @@ -24,6 +24,14 @@ #include "sp_head.h" #include "sql_trigger.h" +/** + Implement DELETE SQL word. + + @note Like implementations of other DDL/DML in MySQL, this function + relies on the caller to close the thread tables. This is done in the + end of dispatch_command(). +*/ + bool mysql_delete(THD *thd, TABLE_LIST *table_list, COND *conds, SQL_LIST *order, ha_rows limit, ulonglong options, bool reset_auto_increment) @@ -380,17 +388,6 @@ cleanup: } DBUG_ASSERT(transactional_table || !deleted || thd->transaction.stmt.modified_non_trans_table); free_underlaid_joins(thd, select_lex); - if (transactional_table) - { - if (ha_autocommit_or_rollback(thd,error >= 0)) - error=1; - } - - if (thd->lock) - { - mysql_unlock_tables(thd, thd->lock); - thd->lock=0; - } if (error < 0 || (thd->lex->ignore && !thd->is_fatal_error)) { thd->row_count_func= deleted; @@ -751,11 +748,9 @@ void multi_delete::abort() The same if all tables are transactional, regardless of where we are. In all other cases do attempt deletes ... */ - if ((table_being_deleted == delete_tables && - table_being_deleted->table->file->has_transactions()) || - !normal_tables) - ha_rollback_stmt(thd); - else if (do_delete) + if (do_delete && normal_tables && + (table_being_deleted != delete_tables || + !table_being_deleted->table->file->has_transactions())) { /* We have to execute the recorded do_deletes() and write info into the @@ -921,11 +916,6 @@ bool multi_delete::send_eof() if (local_error != 0) error_handled= TRUE; // to force early leave from ::send_error() - /* Commit or rollback the current SQL statement */ - if (transactional_tables) - if (ha_autocommit_or_rollback(thd,local_error > 0)) - local_error=1; - if (!local_error) { thd->row_count_func= deleted; @@ -1055,6 +1045,12 @@ trunc_by_del: error= mysql_delete(thd, table_list, (COND*) 0, (SQL_LIST*) 0, HA_POS_ERROR, LL(0), TRUE); ha_enable_transaction(thd, TRUE); + /* + Safety, in case the engine ignored ha_enable_transaction(FALSE) + above. Also clears thd->transaction.*. + */ + error= ha_autocommit_or_rollback(thd, error); + ha_commit(thd); thd->options= save_options; thd->current_stmt_binlog_row_based= save_binlog_row_based; DBUG_RETURN(error); diff --git a/sql/sql_do.cc b/sql/sql_do.cc index a3eb93f87da..af9fa9444e8 100644 --- a/sql/sql_do.cc +++ b/sql/sql_do.cc @@ -28,7 +28,17 @@ bool mysql_do(THD *thd, List<Item> &values) while ((value = li++)) value->val_int(); free_underlaid_joins(thd, &thd->lex->select_lex); - thd->clear_error(); // DO always is OK + + if (thd->is_error()) + { + /* + Rollback the effect of the statement, since next instruction + will clear the error and the rollback in the end of + dispatch_command() won't work. + */ + ha_autocommit_or_rollback(thd, thd->is_error()); + thd->clear_error(); // DO always is OK + } send_ok(thd); DBUG_RETURN(FALSE); } diff --git a/sql/sql_insert.cc b/sql/sql_insert.cc index c2b1990f6c7..b03e786d1a1 100644 --- a/sql/sql_insert.cc +++ b/sql/sql_insert.cc @@ -541,6 +541,10 @@ bool open_and_lock_for_insert_delayed(THD *thd, TABLE_LIST *table_list) /** INSERT statement implementation + + @note Like implementations of other DDL/DML in MySQL, this function + relies on the caller to close the thread tables. This is done in the + end of dispatch_command(). */ bool mysql_insert(THD *thd,TABLE_LIST *table_list, @@ -893,12 +897,9 @@ bool mysql_insert(THD *thd,TABLE_LIST *table_list, } DBUG_ASSERT(transactional_table || !changed || thd->transaction.stmt.modified_non_trans_table); - if (transactional_table) - error=ha_autocommit_or_rollback(thd,error); - + if (thd->lock) { - mysql_unlock_tables(thd, thd->lock); /* Invalidate the table in the query cache if something changed after unlocking when changes become fisible. @@ -909,7 +910,6 @@ bool mysql_insert(THD *thd,TABLE_LIST *table_list, { query_cache_invalidate3(thd, table_list, 1); } - thd->lock=0; } } thd_proc_info(thd, "end"); @@ -2445,7 +2445,7 @@ err: first call to ha_*_row() instead. Remove code that are used to cover for the case outlined above. */ - ha_rollback_stmt(thd); + ha_autocommit_or_rollback(thd, 1); #ifndef __WIN__ end: @@ -3139,18 +3139,6 @@ bool select_insert::send_eof() thd->query, thd->query_length, trans_table, FALSE, killed_status); } - /* - We will call ha_autocommit_or_rollback() also for - non-transactional tables under row-based replication: there might - be events in the binary logs transaction, and we need to write - them to the binary log. - */ - if (trans_table || thd->current_stmt_binlog_row_based) - { - int error2= ha_autocommit_or_rollback(thd, error); - if (error2 && !error) - error= error2; - } table->file->ha_release_auto_increment(); if (error) @@ -3228,7 +3216,6 @@ void select_insert::abort() { table->file->ha_release_auto_increment(); } - ha_rollback_stmt(thd); DBUG_VOID_RETURN; } @@ -3667,7 +3654,10 @@ bool select_create::send_eof() nevertheless. */ if (!table->s->tmp_table) - ha_commit(thd); // Can fail, but we proceed anyway + { + ha_autocommit_or_rollback(thd, 0); + end_active_trans(thd); + } table->file->extra(HA_EXTRA_NO_IGNORE_DUP_KEY); table->file->extra(HA_EXTRA_WRITE_CANNOT_REPLACE); @@ -3691,12 +3681,9 @@ void select_create::abort() by removing the table, even for non-transactional tables. */ tmp_disable_binlog(thd); - select_insert::abort(); - reenable_binlog(thd); - /* - We roll back the statement, including truncating the transaction - cache of the binary log, if the statement failed. + In select_insert::abort() we roll back the statement, including + truncating the transaction cache of the binary log. We roll back the statement prior to deleting the table and prior to releasing the lock on the table, since there might be potential @@ -3707,8 +3694,9 @@ void select_create::abort() of the table succeeded or not, since we need to reset the binary log state. */ - if (thd->current_stmt_binlog_row_based) - ha_rollback_stmt(thd); + select_insert::abort(); + reenable_binlog(thd); + if (m_plock) { diff --git a/sql/sql_load.cc b/sql/sql_load.cc index c96fbb80b0c..d50f6c19a33 100644 --- a/sql/sql_load.cc +++ b/sql/sql_load.cc @@ -470,9 +470,6 @@ bool mysql_load(THD *thd,sql_exchange *ex,TABLE_LIST *table_list, } } #endif /*!EMBEDDED_LIBRARY*/ - if (transactional_table) - ha_autocommit_or_rollback(thd,error); - error= -1; // Error on read goto err; } @@ -510,8 +507,6 @@ bool mysql_load(THD *thd,sql_exchange *ex,TABLE_LIST *table_list, } } #endif /*!EMBEDDED_LIBRARY*/ - if (transactional_table) - error=ha_autocommit_or_rollback(thd,error); /* ok to client sent only after binlog write and engine commit */ send_ok(thd, info.copied + info.deleted, 0L, name); @@ -519,11 +514,6 @@ err: DBUG_ASSERT(transactional_table || !(info.copied || info.deleted) || thd->transaction.stmt.modified_non_trans_table); table->file->ha_release_auto_increment(); - if (thd->lock) - { - mysql_unlock_tables(thd, thd->lock); - thd->lock=0; - } table->auto_increment_field_not_null= FALSE; thd->abort_on_warning= 0; DBUG_RETURN(error); diff --git a/sql/sql_parse.cc b/sql/sql_parse.cc index 221c156fe6b..9b69dd60c45 100644 --- a/sql/sql_parse.cc +++ b/sql/sql_parse.cc @@ -1465,21 +1465,13 @@ bool dispatch_command(enum enum_server_command command, THD *thd, break; } - thd_proc_info(thd, "closing tables"); - /* Free tables */ - close_thread_tables(thd); + /* If commit fails, we should be able to reset the OK status. */ + thd->main_da.can_overwrite_status= TRUE; + ha_autocommit_or_rollback(thd, thd->is_error()); + thd->main_da.can_overwrite_status= FALSE; + + thd->transaction.stmt.reset(); - /* - assume handlers auto-commit (if some doesn't - transaction handling - in MySQL should be redesigned to support it; it's a big change, - and it's not worth it - better to commit explicitly only writing - transactions, read-only ones should better take care of themselves. - saves some work in 2pc too) - see also sql_base.cc - close_thread_tables() - */ - bzero(&thd->transaction.stmt, sizeof(thd->transaction.stmt)); - if (!thd->active_transaction()) - thd->transaction.xid_state.xid.null(); /* report error issued during command execution */ if (thd->killed_errno()) @@ -1496,6 +1488,10 @@ bool dispatch_command(enum enum_server_command command, THD *thd, net_end_statement(thd); query_cache_end_of_result(thd); + thd->proc_info= "closing tables"; + /* Free tables */ + close_thread_tables(thd); + log_slow_statement(thd); thd_proc_info(thd, "cleaning up"); @@ -3011,10 +3007,8 @@ end_with_restore_list: /* INSERT ... SELECT should invalidate only the very first table */ TABLE_LIST *save_table= first_table->next_local; first_table->next_local= 0; - mysql_unlock_tables(thd, thd->lock); query_cache_invalidate3(thd, first_table, 1); first_table->next_local= save_table; - thd->lock=0; } delete sel_result; } @@ -3985,7 +3979,6 @@ end_with_restore_list: push_warning(thd, MYSQL_ERROR::WARN_LEVEL_WARN, ER_PROC_AUTO_GRANT_FAIL, ER(ER_PROC_AUTO_GRANT_FAIL)); - close_thread_tables(thd); } #endif break; diff --git a/sql/sql_partition.cc b/sql/sql_partition.cc index eabf4526f7b..cf0aeee938b 100644 --- a/sql/sql_partition.cc +++ b/sql/sql_partition.cc @@ -3968,31 +3968,35 @@ static int fast_end_partition(THD *thd, ulonglong copied, bool written_bin_log) { int error; + char tmp_name[80]; DBUG_ENTER("fast_end_partition"); thd->proc_info="end"; + if (!is_empty) query_cache_invalidate3(thd, table_list, 0); - error= ha_commit_stmt(thd); - if (ha_commit(thd)) + + error= ha_autocommit_or_rollback(thd, 0); + if (end_active_trans(thd)) error= 1; - if (!error || is_empty) - { - char tmp_name[80]; - if ((!is_empty) && (!written_bin_log) && - (!thd->lex->no_write_to_binlog)) - write_bin_log(thd, FALSE, thd->query, thd->query_length); - close_thread_tables(thd); - my_snprintf(tmp_name, sizeof(tmp_name), ER(ER_INSERT_INFO), - (ulong) (copied + deleted), - (ulong) deleted, - (ulong) 0); - send_ok(thd, (ha_rows) (copied+deleted),0L,tmp_name); - DBUG_RETURN(FALSE); + + if (error) + { + /* If error during commit, no need to rollback, it's done. */ + table->file->print_error(error, MYF(0)); + DBUG_RETURN(TRUE); } - table->file->print_error(error, MYF(0)); - close_thread_tables(thd); - DBUG_RETURN(TRUE); + + if ((!is_empty) && (!written_bin_log) && + (!thd->lex->no_write_to_binlog)) + write_bin_log(thd, FALSE, thd->query, thd->query_length); + + my_snprintf(tmp_name, sizeof(tmp_name), ER(ER_INSERT_INFO), + (ulong) (copied + deleted), + (ulong) deleted, + (ulong) 0); + send_ok(thd, (ha_rows) (copied+deleted),0L, tmp_name); + DBUG_RETURN(FALSE); } diff --git a/sql/sql_table.cc b/sql/sql_table.cc index 505bcd1b421..61f8ff342b1 100644 --- a/sql/sql_table.cc +++ b/sql/sql_table.cc @@ -4131,6 +4131,7 @@ static bool mysql_admin_table(THD* thd, TABLE_LIST* tables, switch ((*prepare_func)(thd, table, check_opt)) { case 1: // error, message written to net ha_autocommit_or_rollback(thd, 1); + end_trans(thd, ROLLBACK); close_thread_tables(thd); DBUG_PRINT("admin", ("simple error, admin next table")); continue; @@ -4189,6 +4190,7 @@ static bool mysql_admin_table(THD* thd, TABLE_LIST* tables, table_name); protocol->store(buff, length, system_charset_info); ha_autocommit_or_rollback(thd, 0); + end_trans(thd, COMMIT); close_thread_tables(thd); lex->reset_query_tables_list(FALSE); table->table=0; // For query cache @@ -4461,6 +4463,7 @@ send_result_message: } } ha_autocommit_or_rollback(thd, 0); + end_trans(thd, COMMIT); close_thread_tables(thd); table->table=0; // For query cache if (protocol->write()) @@ -4470,8 +4473,9 @@ send_result_message: send_eof(thd); DBUG_RETURN(FALSE); - err: +err: ha_autocommit_or_rollback(thd, 1); + end_trans(thd, ROLLBACK); close_thread_tables(thd); // Shouldn't be needed if (table) table->table=0; @@ -4994,8 +4998,8 @@ mysql_discard_or_import_tablespace(THD *thd, query_cache_invalidate3(thd, table_list, 0); /* The ALTER TABLE is always in its own transaction */ - error = ha_commit_stmt(thd); - if (ha_commit(thd)) + error = ha_autocommit_or_rollback(thd, 0); + if (end_active_trans(thd)) error=1; if (error) goto err; @@ -5003,7 +5007,6 @@ mysql_discard_or_import_tablespace(THD *thd, err: ha_autocommit_or_rollback(thd, error); - close_thread_tables(thd); thd->tablespace_op=FALSE; if (error == 0) @@ -6526,8 +6529,8 @@ view_err: VOID(pthread_mutex_unlock(&LOCK_open)); alter_table_manage_keys(table, table->file->indexes_are_disabled(), alter_info->keys_onoff); - error= ha_commit_stmt(thd); - if (ha_commit(thd)) + error= ha_autocommit_or_rollback(thd, 0); + if (end_active_trans(thd)) error= 1; } thd->count_cuted_fields= CHECK_FIELD_IGNORE; @@ -6615,7 +6618,7 @@ view_err: /* Need to commit before a table is unlocked (NDB requirement). */ DBUG_PRINT("info", ("Committing before unlocking table")); - if (ha_commit_stmt(thd) || ha_commit(thd)) + if (ha_autocommit_or_rollback(thd, 0) || end_active_trans(thd)) goto err1; committed= 1; } @@ -7116,9 +7119,9 @@ copy_data_between_tables(TABLE *from,TABLE *to, Ensure that the new table is saved properly to disk so that we can do a rename */ - if (ha_commit_stmt(thd)) + if (ha_autocommit_or_rollback(thd, 0)) error=1; - if (ha_commit(thd)) + if (end_active_trans(thd)) error=1; err: diff --git a/sql/sql_udf.cc b/sql/sql_udf.cc index 30e8829d764..ae3f43eba5b 100644 --- a/sql/sql_udf.cc +++ b/sql/sql_udf.cc @@ -382,6 +382,14 @@ static udf_func *add_udf(LEX_STRING *name, Item_result ret, char *dl, } +/** + Create a user defined function. + + @note Like implementations of other DDL/DML in MySQL, this function + relies on the caller to close the thread tables. This is done in the + end of dispatch_command(). +*/ + int mysql_create_function(THD *thd,udf_func *udf) { int error; @@ -489,7 +497,6 @@ int mysql_create_function(THD *thd,udf_func *udf) table->field[3]->store((longlong) u_d->type, TRUE); error = table->file->ha_write_row(table->record[0]); - close_thread_tables(thd); if (error) { my_error(ER_ERROR_ON_WRITE, MYF(0), "mysql.func", error); diff --git a/sql/sql_update.cc b/sql/sql_update.cc index 4d075e3308d..b1485389db3 100644 --- a/sql/sql_update.cc +++ b/sql/sql_update.cc @@ -803,17 +803,6 @@ int mysql_update(THD *thd, } DBUG_ASSERT(transactional_table || !updated || thd->transaction.stmt.modified_non_trans_table); free_underlaid_joins(thd, select_lex); - if (transactional_table) - { - if (ha_autocommit_or_rollback(thd, error >= 0)) - error=1; - } - - if (thd->lock) - { - mysql_unlock_tables(thd, thd->lock); - thd->lock=0; - } /* If LAST_INSERT_ID(X) was used, report X */ id= thd->arg_of_last_insert_id_function ? @@ -1716,13 +1705,8 @@ void multi_update::abort() If not attempt to do remaining updates. */ - if (trans_safe) + if (! trans_safe) { - DBUG_ASSERT(transactional_tables); - (void) ha_autocommit_or_rollback(thd, 1); - } - else - { DBUG_ASSERT(thd->transaction.stmt.modified_non_trans_table); if (do_update && table_count > 1) { @@ -1754,11 +1738,6 @@ void multi_update::abort() thd->transaction.all.modified_non_trans_table= TRUE; } DBUG_ASSERT(trans_safe || !updated || thd->transaction.stmt.modified_non_trans_table); - - if (transactional_tables) - { - (void) ha_autocommit_or_rollback(thd, 1); - } } @@ -1996,12 +1975,6 @@ bool multi_update::send_eof() if (local_error != 0) error_handled= TRUE; // to force early leave from ::send_error() - if (transactional_tables) - { - if (ha_autocommit_or_rollback(thd, local_error != 0)) - local_error=1; - } - if (local_error > 0) // if the above log write did not fail ... { /* Safety: If we haven't got an error before (can happen in do_updates) */ |