/* Copyright 2000-2008 MySQL AB, 2008-2009 Sun Microsystems, Inc. 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 */ /** @file Locking functions for mysql. Because of the new concurrent inserts, we must first get external locks before getting internal locks. If we do it in the other order, the status information is not up to date when called from the lock handler. GENERAL DESCRIPTION OF LOCKING When not using LOCK TABLES: - For each SQL statement mysql_lock_tables() is called for all involved tables. - mysql_lock_tables() will call table_handler->external_lock(thd,locktype) for each table. This is followed by a call to thr_multi_lock() for all tables. - When statement is done, we call mysql_unlock_tables(). This will call thr_multi_unlock() followed by table_handler->external_lock(thd, F_UNLCK) for each table. - Note that mysql_unlock_tables() may be called several times as MySQL in some cases can free some tables earlier than others. - The above is true both for normal and temporary tables. - Temporary non transactional tables are never passed to thr_multi_lock() and we never call external_lock(thd, F_UNLOCK) on these. When using LOCK TABLES: - LOCK TABLE will call mysql_lock_tables() for all tables. mysql_lock_tables() will call table_handler->external_lock(thd,locktype) for each table. This is followed by a call to thr_multi_lock() for all tables. - For each statement, we will call table_handler->start_stmt(THD) to inform the table handler that we are using the table. The tables used can only be tables used in LOCK TABLES or a temporary table. - When statement is done, we will call ha_commit_stmt(thd); - When calling UNLOCK TABLES we call mysql_unlock_tables() for all tables used in LOCK TABLES If table_handler->external_lock(thd, locktype) fails, we call table_handler->external_lock(thd, F_UNLCK) for each table that was locked, excluding one that caused failure. That means handler must cleanup itself in case external_lock() fails. @todo Change to use my_malloc() ONLY when using LOCK TABLES command or when we are forced to use mysql_lock_merge. */ #include "sql_priv.h" #include "debug_sync.h" #include "unireg.h" // REQUIRED: for other includes #include "lock.h" #include "sql_base.h" // close_tables_for_reopen #include "sql_parse.h" // is_log_table_write_query #include "sql_acl.h" // SUPER_ACL #include #include /** @defgroup Locking Locking @{ */ extern HASH open_cache; /* flags for get_lock_data */ #define GET_LOCK_UNLOCK 1 #define GET_LOCK_STORE_LOCKS 2 static MYSQL_LOCK *get_lock_data(THD *thd, TABLE **table_ptr, uint count, uint flags); static int lock_external(THD *thd, TABLE **table,uint count); static int unlock_external(THD *thd, TABLE **table,uint count); static void print_lock_error(int error, const char *); /* Map the return value of thr_lock to an error from errmsg.txt */ static int thr_lock_errno_to_mysql[]= { 0, ER_LOCK_ABORTED, ER_LOCK_WAIT_TIMEOUT, ER_LOCK_DEADLOCK }; /** Perform semantic checks for mysql_lock_tables. @param thd The current thread @param tables The tables to lock @param count The number of tables to lock @param flags Lock flags @return 0 if all the check passed, non zero if a check failed. */ static int lock_tables_check(THD *thd, TABLE **tables, uint count, uint flags) { uint system_count, i; bool is_superuser, log_table_write_query; DBUG_ENTER("lock_tables_check"); system_count= 0; is_superuser= thd->security_ctx->master_access & SUPER_ACL; log_table_write_query= (is_log_table_write_query(thd->lex->sql_command) || ((flags & MYSQL_LOCK_LOG_TABLE) != 0)); for (i=0 ; is->table_category != TABLE_UNKNOWN_CATEGORY); /* Table I/O to performance schema tables is performed only internally by the server implementation. When a user is requesting a lock, the following constraints are enforced: */ if (t->s->require_write_privileges() && ! log_table_write_query) { /* A user should not be able to prevent writes, or hold any type of lock in a session, since this would be a DOS attack. */ if ((t->reginfo.lock_type >= TL_READ_NO_INSERT) || (thd->lex->sql_command == SQLCOM_LOCK_TABLES)) { my_error(ER_CANT_LOCK_LOG_TABLE, MYF(0)); DBUG_RETURN(1); } } if (t->reginfo.lock_type >= TL_WRITE_ALLOW_WRITE) { if (t->s->table_category == TABLE_CATEGORY_SYSTEM) system_count++; if (t->db_stat & HA_READ_ONLY) { my_error(ER_OPEN_AS_READONLY, MYF(0), t->alias); DBUG_RETURN(1); } } /* If we are going to lock a non-temporary table we must own metadata lock of appropriate type on it (I.e. for table to be locked for write we must own metadata lock of MDL_SHARED_WRITE or stronger type. For table to be locked for read we must own metadata lock of MDL_SHARED_READ or stronger type). The only exception are HANDLER statements which are allowed to lock table for read while having only MDL_SHARED lock on it. */ DBUG_ASSERT(t->s->tmp_table || thd->mdl_context.is_lock_owner(MDL_key::TABLE, t->s->db.str, t->s->table_name.str, t->reginfo.lock_type >= TL_WRITE_ALLOW_WRITE ? MDL_SHARED_WRITE : MDL_SHARED_READ) || (t->open_by_handler && thd->mdl_context.is_lock_owner(MDL_key::TABLE, t->s->db.str, t->s->table_name.str, MDL_SHARED))); /* Prevent modifications to base tables if READ_ONLY is activated. In any case, read only does not apply to temporary tables. */ if (!(flags & MYSQL_LOCK_IGNORE_GLOBAL_READ_ONLY) && !t->s->tmp_table) { if (t->reginfo.lock_type >= TL_WRITE_ALLOW_WRITE && !is_superuser && opt_readonly && !thd->slave_thread) { my_error(ER_OPTION_PREVENTS_STATEMENT, MYF(0), "--read-only"); DBUG_RETURN(1); } } } /* Locking of system tables is restricted: locking a mix of system and non-system tables in the same lock is prohibited, to prevent contention. */ if ((system_count > 0) && (system_count < count)) { my_error(ER_WRONG_LOCK_OF_SYSTEM_TABLE, MYF(0)); DBUG_RETURN(1); } DBUG_RETURN(0); } /** Reset lock type in lock data @param mysql_lock Lock structures to reset. @note After a locking error we want to quit the locking of the table(s). The test case in the bug report for Bug #18544 has the following cases: 1. Locking error in lock_external() due to InnoDB timeout. 2. Locking error in get_lock_data() due to missing write permission. 3. Locking error in wait_if_global_read_lock() due to lock conflict. @note In all these cases we have already set the lock type into the lock data of the open table(s). If the table(s) are in the open table cache, they could be reused with the non-zero lock type set. This could lead to ignoring a different lock type with the next lock. @note Clear the lock type of all lock data. This ensures that the next lock request will set its lock type properly. */ static void reset_lock_data(MYSQL_LOCK *sql_lock) { THR_LOCK_DATA **ldata, **ldata_end; DBUG_ENTER("reset_lock_data"); /* Clear the lock type of all lock data to avoid reusage. */ for (ldata= sql_lock->locks, ldata_end= ldata + sql_lock->lock_count; ldata < ldata_end; ldata++) { /* Reset lock type. */ (*ldata)->type= TL_UNLOCK; } DBUG_VOID_RETURN; } /** Reset lock type in lock data and free. @param mysql_lock Lock structures to reset. */ static void reset_lock_data_and_free(MYSQL_LOCK **mysql_lock) { reset_lock_data(*mysql_lock); my_free(*mysql_lock, MYF(0)); *mysql_lock= 0; } /** Lock tables. @param thd The current thread. @param tables An array of pointers to the tables to lock. @param count The number of tables to lock. @param flags Options: MYSQL_LOCK_IGNORE_GLOBAL_READ_ONLY Ignore SET GLOBAL READ_ONLY MYSQL_LOCK_IGNORE_TIMEOUT Use maximum timeout value. @retval A lock structure pointer on success. @retval NULL if an error or if wait on a lock was killed. */ MYSQL_LOCK *mysql_lock_tables(THD *thd, TABLE **tables, uint count, uint flags) { int rc; MYSQL_LOCK *sql_lock; ulong timeout= (flags & MYSQL_LOCK_IGNORE_TIMEOUT) ? LONG_TIMEOUT : thd->variables.lock_wait_timeout; DBUG_ENTER("mysql_lock_tables"); if (lock_tables_check(thd, tables, count, flags)) DBUG_RETURN(NULL); if (! (sql_lock= get_lock_data(thd, tables, count, GET_LOCK_STORE_LOCKS))) DBUG_RETURN(NULL); thd_proc_info(thd, "System lock"); DBUG_PRINT("info", ("thd->proc_info %s", thd->proc_info)); if (sql_lock->table_count && lock_external(thd, sql_lock->table, sql_lock->table_count)) { /* Clear the lock type of all lock data to avoid reusage. */ reset_lock_data_and_free(&sql_lock); goto end; } /* Copy the lock data array. thr_multi_lock() reorders its contents. */ memcpy(sql_lock->locks + sql_lock->lock_count, sql_lock->locks, sql_lock->lock_count * sizeof(*sql_lock->locks)); /* Lock on the copied half of the lock data array. */ rc= thr_lock_errno_to_mysql[(int) thr_multi_lock(sql_lock->locks + sql_lock->lock_count, sql_lock->lock_count, thd->lock_id, timeout)]; if (rc) { if (sql_lock->table_count) (void) unlock_external(thd, sql_lock->table, sql_lock->table_count); reset_lock_data_and_free(&sql_lock); if (! thd->killed) my_error(rc, MYF(0)); } end: thd_proc_info(thd, 0); if (thd->killed) { thd->send_kill_message(); if (sql_lock) { mysql_unlock_tables(thd, sql_lock); sql_lock= 0; } } thd->set_time_after_lock(); DBUG_RETURN(sql_lock); } static int lock_external(THD *thd, TABLE **tables, uint count) { reg1 uint i; int lock_type,error; DBUG_ENTER("lock_external"); DBUG_PRINT("info", ("count %d", count)); for (i=1 ; i <= count ; i++, tables++) { DBUG_ASSERT((*tables)->reginfo.lock_type >= TL_READ); lock_type=F_WRLCK; /* Lock exclusive */ if ((*tables)->db_stat & HA_READ_ONLY || ((*tables)->reginfo.lock_type >= TL_READ && (*tables)->reginfo.lock_type <= TL_READ_NO_INSERT)) lock_type=F_RDLCK; if ((error=(*tables)->file->ha_external_lock(thd,lock_type))) { print_lock_error(error, (*tables)->file->table_type()); while (--i) { tables--; (*tables)->file->ha_external_lock(thd, F_UNLCK); (*tables)->current_lock=F_UNLCK; } DBUG_RETURN(error); } else { (*tables)->db_stat &= ~ HA_BLOCK_LOCK; (*tables)->current_lock= lock_type; } } DBUG_RETURN(0); } void mysql_unlock_tables(THD *thd, MYSQL_LOCK *sql_lock) { DBUG_ENTER("mysql_unlock_tables"); if (sql_lock->lock_count) thr_multi_unlock(sql_lock->locks,sql_lock->lock_count); if (sql_lock->table_count) (void) unlock_external(thd,sql_lock->table,sql_lock->table_count); my_free((uchar*) sql_lock,MYF(0)); DBUG_VOID_RETURN; } /** Unlock some of the tables locked by mysql_lock_tables. This will work even if get_lock_data fails (next unlock will free all) */ void mysql_unlock_some_tables(THD *thd, TABLE **table,uint count) { MYSQL_LOCK *sql_lock; if ((sql_lock= get_lock_data(thd, table, count, GET_LOCK_UNLOCK))) mysql_unlock_tables(thd, sql_lock); } /** unlock all tables locked for read. */ void mysql_unlock_read_tables(THD *thd, MYSQL_LOCK *sql_lock) { uint i,found; DBUG_ENTER("mysql_unlock_read_tables"); /* Move all write locks first */ THR_LOCK_DATA **lock=sql_lock->locks; for (i=found=0 ; i < sql_lock->lock_count ; i++) { if (sql_lock->locks[i]->type >= TL_WRITE_ALLOW_READ) { swap_variables(THR_LOCK_DATA *, *lock, sql_lock->locks[i]); lock++; found++; } } /* unlock the read locked tables */ if (i != found) { thr_multi_unlock(lock,i-found); sql_lock->lock_count= found; } /* Then do the same for the external locks */ /* Move all write locked tables first */ TABLE **table=sql_lock->table; for (i=found=0 ; i < sql_lock->table_count ; i++) { DBUG_ASSERT(sql_lock->table[i]->lock_position == i); if ((uint) sql_lock->table[i]->reginfo.lock_type >= TL_WRITE_ALLOW_READ) { swap_variables(TABLE *, *table, sql_lock->table[i]); table++; found++; } } /* Unlock all read locked tables */ if (i != found) { (void) unlock_external(thd,table,i-found); sql_lock->table_count=found; } /* Fix the lock positions in TABLE */ table= sql_lock->table; found= 0; for (i= 0; i < sql_lock->table_count; i++) { TABLE *tbl= *table; tbl->lock_position= (uint) (table - sql_lock->table); tbl->lock_data_start= found; found+= tbl->lock_count; table++; } DBUG_VOID_RETURN; } /** Try to find the table in the list of locked tables. In case of success, unlock the table and remove it from this list. If a table has more than one lock instance, removes them all. @param thd thread context @param locked list of locked tables @param table the table to unlock */ void mysql_lock_remove(THD *thd, MYSQL_LOCK *locked,TABLE *table) { if (locked) { reg1 uint i; for (i=0; i < locked->table_count; i++) { if (locked->table[i] == table) { uint j, removed_locks, old_tables; TABLE *tbl; uint lock_data_end; DBUG_ASSERT(table->lock_position == i); /* Unlock the table. */ mysql_unlock_some_tables(thd, &table, /* table count */ 1); /* Decrement table_count in advance, making below expressions easier */ old_tables= --locked->table_count; /* The table has 'removed_locks' lock data elements in locked->locks */ removed_locks= table->lock_count; /* Move down all table pointers above 'i'. */ bmove((char*) (locked->table+i), (char*) (locked->table+i+1), (old_tables - i) * sizeof(TABLE*)); lock_data_end= table->lock_data_start + table->lock_count; /* Move down all lock data pointers above 'table->lock_data_end-1' */ bmove((char*) (locked->locks + table->lock_data_start), (char*) (locked->locks + lock_data_end), (locked->lock_count - lock_data_end) * sizeof(THR_LOCK_DATA*)); /* Fix moved table elements. lock_position is the index in the 'locked->table' array, it must be fixed by one. table->lock_data_start is pointer to the lock data for this table in the 'locked->locks' array, they must be fixed by 'removed_locks', the lock data count of the removed table. */ for (j= i ; j < old_tables; j++) { tbl= locked->table[j]; tbl->lock_position--; DBUG_ASSERT(tbl->lock_position == j); tbl->lock_data_start-= removed_locks; } /* Finally adjust lock_count. */ locked->lock_count-= removed_locks; break; } } } } /* Downgrade all locks on a table to new WRITE level from WRITE_ONLY */ void mysql_lock_downgrade_write(THD *thd, TABLE *table, thr_lock_type new_lock_type) { MYSQL_LOCK *locked; if ((locked = get_lock_data(thd, &table, 1, GET_LOCK_UNLOCK))) { for (uint i=0; i < locked->lock_count; i++) thr_downgrade_write_lock(locked->locks[i], new_lock_type); my_free((uchar*) locked,MYF(0)); } } /** Abort all other threads waiting to get lock in table. */ void mysql_lock_abort(THD *thd, TABLE *table, bool upgrade_lock) { MYSQL_LOCK *locked; DBUG_ENTER("mysql_lock_abort"); if ((locked= get_lock_data(thd, &table, 1, GET_LOCK_UNLOCK))) { for (uint i=0; i < locked->lock_count; i++) thr_abort_locks(locked->locks[i]->lock, upgrade_lock); my_free((uchar*) locked,MYF(0)); } DBUG_VOID_RETURN; } /** Abort one thread / table combination. @param thd Thread handler @param table Table that should be removed from lock queue @retval 0 Table was not locked by another thread @retval 1 Table was locked by at least one other thread */ bool mysql_lock_abort_for_thread(THD *thd, TABLE *table) { MYSQL_LOCK *locked; bool result= FALSE; DBUG_ENTER("mysql_lock_abort_for_thread"); if ((locked= get_lock_data(thd, &table, 1, GET_LOCK_UNLOCK))) { for (uint i=0; i < locked->lock_count; i++) { if (thr_abort_locks_for_thread(locked->locks[i]->lock, table->in_use->thread_id)) result= TRUE; } my_free((uchar*) locked,MYF(0)); } DBUG_RETURN(result); } MYSQL_LOCK *mysql_lock_merge(MYSQL_LOCK *a,MYSQL_LOCK *b) { MYSQL_LOCK *sql_lock; TABLE **table, **end_table; DBUG_ENTER("mysql_lock_merge"); if (!(sql_lock= (MYSQL_LOCK*) my_malloc(sizeof(*sql_lock)+ sizeof(THR_LOCK_DATA*)*(a->lock_count+b->lock_count)+ sizeof(TABLE*)*(a->table_count+b->table_count),MYF(MY_WME)))) DBUG_RETURN(0); // Fatal error sql_lock->lock_count=a->lock_count+b->lock_count; sql_lock->table_count=a->table_count+b->table_count; sql_lock->locks=(THR_LOCK_DATA**) (sql_lock+1); sql_lock->table=(TABLE**) (sql_lock->locks+sql_lock->lock_count); memcpy(sql_lock->locks,a->locks,a->lock_count*sizeof(*a->locks)); memcpy(sql_lock->locks+a->lock_count,b->locks, b->lock_count*sizeof(*b->locks)); memcpy(sql_lock->table,a->table,a->table_count*sizeof(*a->table)); memcpy(sql_lock->table+a->table_count,b->table, b->table_count*sizeof(*b->table)); /* Now adjust lock_position and lock_data_start for all objects that was moved in 'b' (as there is now all objects in 'a' before these). */ for (table= sql_lock->table + a->table_count, end_table= table + b->table_count; table < end_table; table++) { (*table)->lock_position+= a->table_count; (*table)->lock_data_start+= a->lock_count; } /* Delete old, not needed locks */ my_free((uchar*) a,MYF(0)); my_free((uchar*) b,MYF(0)); thr_lock_merge_status(sql_lock->locks, sql_lock->lock_count); DBUG_RETURN(sql_lock); } /** Find duplicate lock in tables. Temporary tables are ignored here like they are ignored in get_lock_data(). If we allow two opens on temporary tables later, both functions should be checked. @param thd The current thread. @param needle The table to check for duplicate lock. @param haystack The list of tables to search for the dup lock. @note This is mainly meant for MERGE tables in INSERT ... SELECT situations. The 'real', underlying tables can be found only after the MERGE tables are opened. This function assumes that the tables are already locked. @retval NULL No duplicate lock found. @retval !NULL First table from 'haystack' that matches a lock on 'needle'. */ TABLE_LIST *mysql_lock_have_duplicate(THD *thd, TABLE_LIST *needle, TABLE_LIST *haystack) { MYSQL_LOCK *mylock; TABLE **lock_tables; TABLE *table; TABLE *table2; THR_LOCK_DATA **lock_locks; THR_LOCK_DATA **table_lock_data; THR_LOCK_DATA **end_data; THR_LOCK_DATA **lock_data2; THR_LOCK_DATA **end_data2; DBUG_ENTER("mysql_lock_have_duplicate"); /* Table may not be defined for derived or view tables. Table may not be part of a lock for delayed operations. */ if (! (table= needle->table) || ! table->lock_count) goto end; /* A temporary table does not have locks. */ if (table->s->tmp_table == NON_TRANSACTIONAL_TMP_TABLE) goto end; /* Get command lock or LOCK TABLES lock. Maybe empty for INSERT DELAYED. */ if (! (mylock= thd->lock)) goto end; /* If we have less than two tables, we cannot have duplicates. */ if (mylock->table_count < 2) goto end; lock_locks= mylock->locks; lock_tables= mylock->table; /* Prepare table related variables that don't change in loop. */ DBUG_ASSERT((table->lock_position < mylock->table_count) && (table == lock_tables[table->lock_position])); table_lock_data= lock_locks + table->lock_data_start; end_data= table_lock_data + table->lock_count; for (; haystack; haystack= haystack->next_global) { if (haystack->placeholder()) continue; table2= haystack->table; if (table2->s->tmp_table == NON_TRANSACTIONAL_TMP_TABLE) continue; /* All tables in list must be in lock. */ DBUG_ASSERT((table2->lock_position < mylock->table_count) && (table2 == lock_tables[table2->lock_position])); for (lock_data2= lock_locks + table2->lock_data_start, end_data2= lock_data2 + table2->lock_count; lock_data2 < end_data2; lock_data2++) { THR_LOCK_DATA **lock_data; THR_LOCK *lock2= (*lock_data2)->lock; for (lock_data= table_lock_data; lock_data < end_data; lock_data++) { if ((*lock_data)->lock == lock2) { DBUG_PRINT("info", ("haystack match: '%s'", haystack->table_name)); DBUG_RETURN(haystack); } } } } end: DBUG_PRINT("info", ("no duplicate found")); DBUG_RETURN(NULL); } /** Unlock a set of external. */ static int unlock_external(THD *thd, TABLE **table,uint count) { int error,error_code; DBUG_ENTER("unlock_external"); error_code=0; do { if ((*table)->current_lock != F_UNLCK) { (*table)->current_lock = F_UNLCK; if ((error=(*table)->file->ha_external_lock(thd, F_UNLCK))) { error_code=error; print_lock_error(error_code, (*table)->file->table_type()); } } table++; } while (--count); DBUG_RETURN(error_code); } /** Get lock structures from table structs and initialize locks. @param thd Thread handler @param table_ptr Pointer to tables that should be locks @param flags One of: - GET_LOCK_UNLOCK : If we should send TL_IGNORE to store lock - GET_LOCK_STORE_LOCKS : Store lock info in TABLE */ static MYSQL_LOCK *get_lock_data(THD *thd, TABLE **table_ptr, uint count, uint flags) { uint i,tables,lock_count; MYSQL_LOCK *sql_lock; THR_LOCK_DATA **locks, **locks_buf, **locks_start; TABLE **to, **table_buf; DBUG_ENTER("get_lock_data"); DBUG_ASSERT((flags == GET_LOCK_UNLOCK) || (flags == GET_LOCK_STORE_LOCKS)); DBUG_PRINT("info", ("count %d", count)); for (i=tables=lock_count=0 ; i < count ; i++) { TABLE *t= table_ptr[i]; if (t->s->tmp_table != NON_TRANSACTIONAL_TMP_TABLE) { tables+= t->file->lock_count(); lock_count++; } } /* Allocating twice the number of pointers for lock data for use in thr_mulit_lock(). This function reorders the lock data, but cannot update the table values. So the second part of the array is copied from the first part immediately before calling thr_multi_lock(). */ if (!(sql_lock= (MYSQL_LOCK*) my_malloc(sizeof(*sql_lock) + sizeof(THR_LOCK_DATA*) * tables * 2 + sizeof(table_ptr) * lock_count, MYF(0)))) DBUG_RETURN(0); locks= locks_buf= sql_lock->locks= (THR_LOCK_DATA**) (sql_lock + 1); to= table_buf= sql_lock->table= (TABLE**) (locks + tables * 2); sql_lock->table_count=lock_count; for (i=0 ; i < count ; i++) { TABLE *table; enum thr_lock_type lock_type; THR_LOCK_DATA **org_locks = locks; if ((table=table_ptr[i])->s->tmp_table == NON_TRANSACTIONAL_TMP_TABLE) continue; lock_type= table->reginfo.lock_type; DBUG_ASSERT(lock_type != TL_WRITE_DEFAULT && lock_type != TL_READ_DEFAULT); locks_start= locks; locks= table->file->store_lock(thd, locks, (flags & GET_LOCK_UNLOCK) ? TL_IGNORE : lock_type); if (flags & GET_LOCK_STORE_LOCKS) { table->lock_position= (uint) (to - table_buf); table->lock_data_start= (uint) (locks_start - locks_buf); table->lock_count= (uint) (locks - locks_start); } *to++= table; if (locks) for ( ; org_locks != locks ; org_locks++) (*org_locks)->debug_print_param= (void *) table; } /* We do not use 'tables', because there are cases where store_lock() returns less locks than lock_count() claimed. This can happen when a FLUSH TABLES tries to abort locks from a MERGE table of another thread. When that thread has just opened the table, but not yet attached its children, it cannot return the locks. lock_count() always returns the number of locks that an attached table has. This is done to avoid the reverse situation: If lock_count() would return 0 for a non-attached MERGE table, and that table becomes attached between the calls to lock_count() and store_lock(), then we would have allocated too little memory for the lock data. Now we may allocate too much, but better safe than memory overrun. And in the FLUSH case, the memory is released quickly anyway. */ sql_lock->lock_count= locks - locks_buf; DBUG_PRINT("info", ("sql_lock->table_count %d sql_lock->lock_count %d", sql_lock->table_count, sql_lock->lock_count)); DBUG_RETURN(sql_lock); } /***************************************************************************** Lock table based on the name. This is used when we need total access to a closed, not open table *****************************************************************************/ /** Obtain exclusive metadata locks on the list of tables. @param thd Thread handle @param table_list List of tables to lock @note This function assumes that no metadata locks were acquired before calling it. Also it cannot be called while holding LOCK_open mutex. Both these invariants are enforced by asserts in MDL_context::acquire_locks(). @retval FALSE Success. @retval TRUE Failure (OOM or thread was killed). */ bool lock_table_names(THD *thd, TABLE_LIST *table_list) { MDL_request_list mdl_requests; MDL_request global_request; TABLE_LIST *lock_table; global_request.init(MDL_key::GLOBAL, "", "", MDL_INTENTION_EXCLUSIVE); for (lock_table= table_list; lock_table; lock_table= lock_table->next_local) { lock_table->mdl_request.init(MDL_key::TABLE, lock_table->db, lock_table->table_name, MDL_EXCLUSIVE); mdl_requests.push_front(&lock_table->mdl_request); } mdl_requests.push_front(&global_request); if (thd->mdl_context.acquire_locks(&mdl_requests, thd->variables.lock_wait_timeout)) return 1; return 0; } /** Release all metadata locks previously obtained by lock_table_names(). @param thd Thread handle. @note Cannot be called while holding LOCK_open mutex. */ void unlock_table_names(THD *thd) { DBUG_ENTER("unlock_table_names"); thd->mdl_context.release_transactional_locks(); DBUG_VOID_RETURN; } /** Obtain an exclusive metadata lock on the stored routine name. @param thd Thread handle. @param is_function Stored routine type (only functions or procedures are name-locked. @param db The schema the routine belongs to. @param name Routine name. This function assumes that no metadata locks were acquired before calling it. Additionally, it cannot be called while holding LOCK_open mutex. Both these invariants are enforced by asserts in MDL_context::acquire_locks(). To avoid deadlocks, we do not try to obtain exclusive metadata locks in LOCK TABLES mode, since in this mode there may be other metadata locks already taken by the current connection, and we must not wait for MDL locks while holding locks. @retval FALSE Success. @retval TRUE Failure: we're in LOCK TABLES mode, or out of memory, or this connection was killed. */ bool lock_routine_name(THD *thd, bool is_function, const char *db, const char *name) { MDL_key::enum_mdl_namespace mdl_type= (is_function ? MDL_key::FUNCTION : MDL_key::PROCEDURE); MDL_request_list mdl_requests; MDL_request global_request; MDL_request mdl_request; if (thd->locked_tables_mode) { my_message(ER_LOCK_OR_ACTIVE_TRANSACTION, ER(ER_LOCK_OR_ACTIVE_TRANSACTION), MYF(0)); return TRUE; } DBUG_ASSERT(name); DEBUG_SYNC(thd, "before_wait_locked_pname"); global_request.init(MDL_key::GLOBAL, "", "", MDL_INTENTION_EXCLUSIVE); mdl_request.init(mdl_type, db, name, MDL_EXCLUSIVE); mdl_requests.push_front(&mdl_request); mdl_requests.push_front(&global_request); if (thd->mdl_context.acquire_locks(&mdl_requests, thd->variables.lock_wait_timeout)) return TRUE; DEBUG_SYNC(thd, "after_wait_locked_pname"); return FALSE; } static void print_lock_error(int error, const char *table) { int textno; DBUG_ENTER("print_lock_error"); switch (error) { case HA_ERR_LOCK_WAIT_TIMEOUT: textno=ER_LOCK_WAIT_TIMEOUT; break; case HA_ERR_READ_ONLY_TRANSACTION: textno=ER_READ_ONLY_TRANSACTION; break; case HA_ERR_LOCK_DEADLOCK: textno=ER_LOCK_DEADLOCK; break; case HA_ERR_WRONG_COMMAND: textno=ER_ILLEGAL_HA; break; default: textno=ER_CANT_LOCK; break; } if ( textno == ER_ILLEGAL_HA ) my_error(textno, MYF(ME_BELL+ME_OLDWIN+ME_WAITTANG), table); else my_error(textno, MYF(ME_BELL+ME_OLDWIN+ME_WAITTANG), error); DBUG_VOID_RETURN; } /**************************************************************************** Handling of global read locks Taking the global read lock is TWO steps (2nd step is optional; without it, COMMIT of existing transactions will be allowed): lock_global_read_lock() THEN make_global_read_lock_block_commit(). The global locks are handled through the global variables: global_read_lock count of threads which have the global read lock (i.e. have completed at least the first step above) global_read_lock_blocks_commit count of threads which have the global read lock and block commits (i.e. are in or have completed the second step above) waiting_for_read_lock count of threads which want to take a global read lock but cannot protect_against_global_read_lock count of threads which have set protection against global read lock. access to them is protected with a mutex LOCK_global_read_lock (XXX: one should never take LOCK_open if LOCK_global_read_lock is taken, otherwise a deadlock may occur. Other mutexes could be a problem too - grep the code for global_read_lock if you want to use any other mutex here) Also one must not hold LOCK_open when calling wait_if_global_read_lock(). When the thread with the global read lock tries to close its tables, it needs to take LOCK_open in close_thread_table(). How blocking of threads by global read lock is achieved: that's advisory. Any piece of code which should be blocked by global read lock must be designed like this: - call to wait_if_global_read_lock(). When this returns 0, no global read lock is owned; if argument abort_on_refresh was 0, none can be obtained. - job - if abort_on_refresh was 0, call to start_waiting_global_read_lock() to allow other threads to get the global read lock. I.e. removal of the protection. (Note: it's a bit like an implementation of rwlock). [ I am sorry to mention some SQL syntaxes below I know I shouldn't but found no better descriptive way ] Why does FLUSH TABLES WITH READ LOCK need to block COMMIT: because it's used to read a non-moving SHOW MASTER STATUS, and a COMMIT writes to the binary log. Why getting the global read lock is two steps and not one. Because FLUSH TABLES WITH READ LOCK needs to insert one other step between the two: flushing tables. So the order is 1) lock_global_read_lock() (prevents any new table write locks, i.e. stalls all new updates) 2) close_cached_tables() (the FLUSH TABLES), which will wait for tables currently opened and being updated to close (so it's possible that there is a moment where all new updates of server are stalled *and* FLUSH TABLES WITH READ LOCK is, too). 3) make_global_read_lock_block_commit(). If we have merged 1) and 3) into 1), we would have had this deadlock: imagine thread 1 and 2, in non-autocommit mode, thread 3, and an InnoDB table t. thd1: SELECT * FROM t FOR UPDATE; thd2: UPDATE t SET a=1; # blocked by row-level locks of thd1 thd3: FLUSH TABLES WITH READ LOCK; # blocked in close_cached_tables() by the table instance of thd2 thd1: COMMIT; # blocked by thd3. thd1 blocks thd2 which blocks thd3 which blocks thd1: deadlock. Note that we need to support that one thread does FLUSH TABLES WITH READ LOCK; and then COMMIT; (that's what innobackup does, for some good reason). So in this exceptional case the COMMIT should not be blocked by the FLUSH TABLES WITH READ LOCK. ****************************************************************************/ volatile uint global_read_lock=0; volatile uint global_read_lock_blocks_commit=0; static volatile uint protect_against_global_read_lock=0; static volatile uint waiting_for_read_lock=0; /** Take global read lock, wait if there is protection against lock. If the global read lock is already taken by this thread, then nothing is done. See also "Handling of global read locks" above. @param thd Reference to thread. @retval False Success, global read lock set, commits are NOT blocked. @retval True Failure, thread was killed. */ bool Global_read_lock::lock_global_read_lock(THD *thd) { DBUG_ENTER("lock_global_read_lock"); if (!m_state) { MDL_request mdl_request; const char *old_message; const char *new_message= "Waiting to get readlock"; (void) mysql_mutex_lock(&LOCK_global_read_lock); #if defined(ENABLED_DEBUG_SYNC) /* The below sync point fires if we have to wait for protect_against_global_read_lock. WARNING: Beware to use WAIT_FOR with this sync point. We hold LOCK_global_read_lock here. Call the sync point before calling enter_cond() as it does use enter_cond() and exit_cond() itself if a WAIT_FOR action is executed in spite of the above warning. Pre-set proc_info so that it is available immediately after the sync point sends a SIGNAL. This makes tests more reliable. */ if (protect_against_global_read_lock) { thd_proc_info(thd, new_message); DEBUG_SYNC(thd, "wait_lock_global_read_lock"); } #endif /* defined(ENABLED_DEBUG_SYNC) */ old_message=thd->enter_cond(&COND_global_read_lock, &LOCK_global_read_lock, new_message); DBUG_PRINT("info", ("waiting_for: %d protect_against: %d", waiting_for_read_lock, protect_against_global_read_lock)); waiting_for_read_lock++; while (protect_against_global_read_lock && !thd->killed) mysql_cond_wait(&COND_global_read_lock, &LOCK_global_read_lock); waiting_for_read_lock--; if (thd->killed) { thd->exit_cond(old_message); DBUG_RETURN(1); } m_state= GRL_ACQUIRED; global_read_lock++; thd->exit_cond(old_message); // this unlocks LOCK_global_read_lock /* When we perform FLUSH TABLES or ALTER TABLE under LOCK TABLES, tables being reopened are protected only by meta-data locks at some point. To avoid sneaking in with our global read lock at this moment we have to take global shared meta data lock. TODO: We should change this code to acquire global shared metadata lock before acquiring global read lock. But in order to do this we have to get rid of all those places in which wait_if_global_read_lock() is called before acquiring metadata locks first. Also long-term we should get rid of redundancy between metadata locks, global read lock and DDL blocker (see WL#4399 and WL#4400). */ DBUG_ASSERT(! thd->mdl_context.is_lock_owner(MDL_key::GLOBAL, "", "", MDL_SHARED)); mdl_request.init(MDL_key::GLOBAL, "", "", MDL_SHARED); if (thd->mdl_context.acquire_lock(&mdl_request, thd->variables.lock_wait_timeout)) { /* Our thread was killed -- return back to initial state. */ mysql_mutex_lock(&LOCK_global_read_lock); if (!(--global_read_lock)) { DBUG_PRINT("signal", ("Broadcasting COND_global_read_lock")); mysql_cond_broadcast(&COND_global_read_lock); } mysql_mutex_unlock(&LOCK_global_read_lock); m_state= GRL_NONE; DBUG_RETURN(1); } thd->mdl_context.move_ticket_after_trans_sentinel(mdl_request.ticket); m_mdl_global_shared_lock= mdl_request.ticket; } /* We DON'T set global_read_lock_blocks_commit now, it will be set after tables are flushed (as the present function serves for FLUSH TABLES WITH READ LOCK only). Doing things in this order is necessary to avoid deadlocks (we must allow COMMIT until all tables are closed; we should not forbid it before, or we can have a 3-thread deadlock if 2 do SELECT FOR UPDATE and one does FLUSH TABLES WITH READ LOCK). */ DBUG_RETURN(0); } /** Unlock global read lock. Commits may or may not be blocked when this function is called. See also "Handling of global read locks" above. @param thd Reference to thread. */ void Global_read_lock::unlock_global_read_lock(THD *thd) { uint tmp; DBUG_ENTER("unlock_global_read_lock"); DBUG_PRINT("info", ("global_read_lock: %u global_read_lock_blocks_commit: %u", global_read_lock, global_read_lock_blocks_commit)); DBUG_ASSERT(m_mdl_global_shared_lock && m_state); thd->mdl_context.release_lock(m_mdl_global_shared_lock); m_mdl_global_shared_lock= NULL; mysql_mutex_lock(&LOCK_global_read_lock); tmp= --global_read_lock; if (m_state == GRL_ACQUIRED_AND_BLOCKS_COMMIT) --global_read_lock_blocks_commit; mysql_mutex_unlock(&LOCK_global_read_lock); /* Send the signal outside the mutex to avoid a context switch */ if (!tmp) { DBUG_PRINT("signal", ("Broadcasting COND_global_read_lock")); mysql_cond_broadcast(&COND_global_read_lock); } m_state= GRL_NONE; DBUG_VOID_RETURN; } /** Wait if the global read lock is set, and optionally seek protection against global read lock. See also "Handling of global read locks" above. @param thd Reference to thread. @param abort_on_refresh If True, abort waiting if a refresh occurs, do NOT seek protection against GRL. If False, wait until the GRL is released and seek protection against GRL. @param is_not_commit If False, called from a commit operation, wait only if commit blocking is also enabled. @retval False Success, protection against global read lock is set (if !abort_on_refresh) @retval True Failure, wait was aborted or thread was killed. */ #define must_wait (global_read_lock && \ (is_not_commit || \ global_read_lock_blocks_commit)) bool Global_read_lock:: wait_if_global_read_lock(THD *thd, bool abort_on_refresh, bool is_not_commit) { const char *UNINIT_VAR(old_message); bool result= 0, need_exit_cond; DBUG_ENTER("wait_if_global_read_lock"); /* If we already have protection against global read lock, just increment the counter. */ if (unlikely(m_protection_count > 0)) { if (!abort_on_refresh) m_protection_count++; DBUG_RETURN(FALSE); } /* Assert that we do not own LOCK_open. If we would own it, other threads could not close their tables. This would make a pretty deadlock. */ mysql_mutex_assert_not_owner(&LOCK_open); mysql_mutex_lock(&LOCK_global_read_lock); if ((need_exit_cond= must_wait)) { if (m_state) // This thread had the read locks { if (is_not_commit) my_message(ER_CANT_UPDATE_WITH_READLOCK, ER(ER_CANT_UPDATE_WITH_READLOCK), MYF(0)); mysql_mutex_unlock(&LOCK_global_read_lock); /* We allow FLUSHer to COMMIT; we assume FLUSHer knows what it does. This allowance is needed to not break existing versions of innobackup which do a BEGIN; INSERT; FLUSH TABLES WITH READ LOCK; COMMIT. */ DBUG_RETURN(is_not_commit); } old_message=thd->enter_cond(&COND_global_read_lock, &LOCK_global_read_lock, "Waiting for release of readlock"); while (must_wait && ! thd->killed && (!abort_on_refresh || thd->version == refresh_version)) { DBUG_PRINT("signal", ("Waiting for COND_global_read_lock")); mysql_cond_wait(&COND_global_read_lock, &LOCK_global_read_lock); DBUG_PRINT("signal", ("Got COND_global_read_lock")); } if (thd->killed) result=1; } if (!abort_on_refresh && !result) { m_protection_count++; protect_against_global_read_lock++; DBUG_PRINT("sql_lock", ("protect_against_global_read_lock incr: %u", protect_against_global_read_lock)); } /* The following is only true in case of a global read locks (which is rare) and if old_message is set */ if (unlikely(need_exit_cond)) thd->exit_cond(old_message); // this unlocks LOCK_global_read_lock else mysql_mutex_unlock(&LOCK_global_read_lock); DBUG_RETURN(result); } /** Release protection against global read lock and restart global read lock waiters. Should only be called if we have protection against global read lock. See also "Handling of global read locks" above. @param thd Reference to thread. */ void Global_read_lock::start_waiting_global_read_lock(THD *thd) { bool tmp; DBUG_ENTER("start_waiting_global_read_lock"); /* Ignore request if we do not have protection against global read lock. (Note that this is a violation of the interface contract, hence the assert). */ DBUG_ASSERT(m_protection_count > 0); if (unlikely(m_protection_count == 0)) DBUG_VOID_RETURN; /* Decrement local read lock protection counter, return if we still have it */ if (unlikely(--m_protection_count > 0)) DBUG_VOID_RETURN; if (unlikely(m_state)) DBUG_VOID_RETURN; mysql_mutex_lock(&LOCK_global_read_lock); DBUG_ASSERT(protect_against_global_read_lock); tmp= (!--protect_against_global_read_lock && (waiting_for_read_lock || global_read_lock_blocks_commit)); mysql_mutex_unlock(&LOCK_global_read_lock); if (tmp) mysql_cond_broadcast(&COND_global_read_lock); DBUG_VOID_RETURN; } /** Make global read lock also block commits. The scenario is: - This thread has the global read lock. - Global read lock blocking of commits is not set. See also "Handling of global read locks" above. @param thd Reference to thread. @retval False Success, global read lock set, commits are blocked. @retval True Failure, thread was killed. */ bool Global_read_lock::make_global_read_lock_block_commit(THD *thd) { bool error; const char *old_message; DBUG_ENTER("make_global_read_lock_block_commit"); /* If we didn't succeed lock_global_read_lock(), or if we already suceeded make_global_read_lock_block_commit(), do nothing. */ if (m_state != GRL_ACQUIRED) DBUG_RETURN(0); mysql_mutex_lock(&LOCK_global_read_lock); /* increment this BEFORE waiting on cond (otherwise race cond) */ global_read_lock_blocks_commit++; /* For testing we set up some blocking, to see if we can be killed */ DBUG_EXECUTE_IF("make_global_read_lock_block_commit_loop", protect_against_global_read_lock++;); old_message= thd->enter_cond(&COND_global_read_lock, &LOCK_global_read_lock, "Waiting for all running commits to finish"); while (protect_against_global_read_lock && !thd->killed) mysql_cond_wait(&COND_global_read_lock, &LOCK_global_read_lock); DBUG_EXECUTE_IF("make_global_read_lock_block_commit_loop", protect_against_global_read_lock--;); if ((error= test(thd->killed))) global_read_lock_blocks_commit--; // undo what we did else m_state= GRL_ACQUIRED_AND_BLOCKS_COMMIT; thd->exit_cond(old_message); // this unlocks LOCK_global_read_lock DBUG_RETURN(error); } /** Broadcast COND_refresh and COND_global_read_lock. Due to a bug in a threading library it could happen that a signal did not reach its target. A condition for this was that the same condition variable was used with different mutexes in mysql_cond_wait(). Some time ago we changed LOCK_open to LOCK_global_read_lock in global read lock handling. So COND_refresh was used with LOCK_open and LOCK_global_read_lock. We did now also change from COND_refresh to COND_global_read_lock in global read lock handling. But now it is necessary to signal both conditions at the same time. @note When signalling COND_global_read_lock within the global read lock handling, it is not necessary to also signal COND_refresh. */ void broadcast_refresh(void) { mysql_cond_broadcast(&COND_refresh); mysql_cond_broadcast(&COND_global_read_lock); } /** @} (end of group Locking) */