Commit 289cc26c authored by unknown's avatar unknown

Bug#21074 Large query_cache freezes mysql server sporadically under heavy load

Invaldating a subset of a sufficiently large query cache can take a long time.
During this time the server is efficiently frozen and no other operation can
be executed. This patch addresses this problem by moving the locks which cause
the freezing and also by temporarily disable the query cache while the 
invalidation takes place.


sql/ha_ndbcluster.cc:
  - mysql_rm_table_part2 has a new parameter to indicate if OPEN_lock mutex 
    protection is needed.
sql/lock.cc:
  - Added function for acquiring table name exclusive locks.
  - Added function for asserting that table name lock is acquired.
sql/mysql_priv.h:
  - Added function for acquiring table name exclusive locks.
  - Added function for asserting that table name lock is acquired.
  - Added parameter to mysql_rm_table_part2 to indicate whether OPEN_lock mutex
    protection is needed or not.
sql/sql_cache.cc:
  - Changed flush_in_progress-flag into a state and added a function, 
    is_flushing() to reflect on this state. A new state was needed to indicate
    that a partial invalidation was in progress.
  - An unused parameter 'under_guard' was removed.
  - The Query_cache mutex structural_guard was pushed down into one
    invalidate_table function to avoid multiple entry points which makes
    maintainens more difficult.
  - Instead of keeping the structural_guard mutex during the entire invalidation
    we set the query cache status state to TABLE_FLUSH_IN_PROGRESS to
    temporarily disable the cache and avoid locking other threads needing the
    Query_cache resource.
sql/sql_cache.h:
  - Changed flush_in_progress-flag into a state and added a function, 
    is_flushing() to reflect on this state. A new state was needed to indicate
    that a partial invalidation was in progress.
  - An unused parameter 'under_guard' was removed.
  - The Query_cache mutex structural_guard was pushed down into one
    invalidate_table function to avoid multiple entry points which makes
    maintainens more difficult.
  - Instead of keeping the structural_guard mutex during the entire invalidation
    we set the query cache status state to TABLE_FLUSH_IN_PROGRESS to
    temporarily disable the cache and avoid locking other threads needing the
    the Query_cache resource.
sql/sql_db.cc:
  - mysql_rm_table_part2_with_lock is redundant and replaced
  with mysql_rm_table_part2.
sql/sql_parse.cc:
  - Function query_cache_invalidate3 isn't protect by a lock and we have a 
    race condition.
  - Moving this function into mysql_rename_tables and make sure it is protected
    by a exclusive table name lock.
sql/sql_rename.cc:
  - Function query_cache_invalidation3 isn't protect by a lock and we have a 
    race condition.
  - Moving this function into mysql_rename_tables and make sure it is protected
    by a exclusive table name lock.
  - Instead of using LOCK_open mutex, which excludes all other threads, the lock
    is changed into exclusive table name locks instead. This prevents us from
    locking the server if a query cache invalidation would take a long time to
    complete.
sql/sql_table.cc:
  - Instead of using LOCK_open mutex, which excludes all other threads, the lock
    is changed into exclusive table name locks instead. This prevents us from
    locking the server if a query cache invalidation would take a long time to
    complete.
  - Added new parameter to mysql_rm_table_part2 to control whether OPEN_lock mutex
    needs to be aquired or not. This is currently needed by the NDB implemenation.
sql/sql_trigger.cc:
  - Table_triggers don't need to be protexted by LOCK_open mutex. This 
    patch cancel this restriction.
  - Refactored comments to doxygen style.
parent 0484d44e
......@@ -6996,7 +6996,6 @@ int ndbcluster_find_files(handlerton *hton, THD *thd,
// Lock mutex before deleting and creating frm files
pthread_mutex_lock(&LOCK_open);
if (!global_read_lock)
{
// Delete old files
......@@ -7010,10 +7009,12 @@ int ndbcluster_find_files(handlerton *hton, THD *thd,
table_list.db= (char*) db;
table_list.alias= table_list.table_name= (char*)file_name;
(void)mysql_rm_table_part2(thd, &table_list,
/* if_exists */ FALSE,
/* drop_temporary */ FALSE,
/* drop_view */ FALSE,
/* dont_log_query*/ TRUE);
FALSE, /* if_exists */
FALSE, /* drop_temporary */
FALSE, /* drop_view */
TRUE, /* dont_log_query*/
FALSE); /* need lock open */
/* Clear error message that is returned when table is deleted */
thd->clear_error();
}
......@@ -7029,7 +7030,7 @@ int ndbcluster_find_files(handlerton *hton, THD *thd,
}
pthread_mutex_unlock(&LOCK_open);
hash_free(&ok_tables);
hash_free(&ndb_tables);
......
......@@ -1027,6 +1027,102 @@ bool lock_table_names(THD *thd, TABLE_LIST *table_list)
}
/**
@brief Lock all tables in list with an exclusive table name lock.
@param thd Thread handle.
@param table_list Names of tables to lock.
@note This function needs to be protected by LOCK_open. If we're
under LOCK TABLES, this function does not work as advertised. Namely,
it does not exclude other threads from using this table and does not
put an exclusive name lock on this table into the table cache.
@see lock_table_names
@see unlock_table_names
@retval TRUE An error occured.
@retval FALSE Name lock successfully acquired.
*/
bool lock_table_names_exclusively(THD *thd, TABLE_LIST *table_list)
{
if (lock_table_names(thd, table_list))
return TRUE;
/*
Upgrade the table name locks from semi-exclusive to exclusive locks.
*/
for (TABLE_LIST *table= table_list; table; table= table->next_global)
{
if (table->table)
table->table->open_placeholder= 1;
}
return FALSE;
}
/**
@brief Test is 'table' is protected by an exclusive name lock.
@param[in] thd The current thread handler
@param[in] table Table container containing the single table to be tested
@note Needs to be protected by LOCK_open mutex.
@return Error status code
@retval TRUE Table is protected
@retval FALSE Table is not protected
*/
bool
is_table_name_exclusively_locked_by_this_thread(THD *thd,
TABLE_LIST *table_list)
{
char key[MAX_DBKEY_LENGTH];
uint key_length;
key_length= create_table_def_key(thd, key, table_list, 0);
return is_table_name_exclusively_locked_by_this_thread(thd, (uchar *)key,
key_length);
}
/**
@brief Test is 'table key' is protected by an exclusive name lock.
@param[in] thd The current thread handler.
@param[in] table Table container containing the single table to be tested.
@note Needs to be protected by LOCK_open mutex
@retval TRUE Table is protected
@retval FALSE Table is not protected
*/
bool
is_table_name_exclusively_locked_by_this_thread(THD *thd, uchar *key,
int key_length)
{
HASH_SEARCH_STATE state;
TABLE *table;
for (table= (TABLE*) hash_first(&open_cache, key,
key_length, &state);
table ;
table= (TABLE*) hash_next(&open_cache, key,
key_length, &state))
{
if (table->in_use == thd &&
table->open_placeholder == 1 &&
table->s->version == 0)
return TRUE;
}
return FALSE;
}
/*
Unlock all tables in list with a name lock
......
......@@ -820,10 +820,8 @@ void mysql_client_binlog_statement(THD *thd);
bool mysql_rm_table(THD *thd,TABLE_LIST *tables, my_bool if_exists,
my_bool drop_temporary);
int mysql_rm_table_part2(THD *thd, TABLE_LIST *tables, bool if_exists,
bool drop_temporary, bool drop_view, bool log_query);
int mysql_rm_table_part2_with_lock(THD *thd, TABLE_LIST *tables,
bool if_exists, bool drop_temporary,
bool log_query);
bool drop_temporary, bool drop_view, bool log_query,
bool need_lock_open);
bool quick_rm_table(handlerton *base,const char *db,
const char *table_name, uint flags);
void close_cached_table(THD *thd, TABLE *table);
......@@ -1799,6 +1797,11 @@ bool wait_for_locked_table_names(THD *thd, TABLE_LIST *table_list);
bool lock_table_names(THD *thd, TABLE_LIST *table_list);
void unlock_table_names(THD *thd, TABLE_LIST *table_list,
TABLE_LIST *last_table);
bool lock_table_names_exclusively(THD *thd, TABLE_LIST *table_list);
bool is_table_name_exclusively_locked_by_this_thread(THD *thd,
TABLE_LIST *table_list);
bool is_table_name_exclusively_locked_by_this_thread(THD *thd, uchar *key,
int key_length);
/* old unireg functions */
......
This diff is collapsed.
......@@ -65,17 +65,44 @@ struct Query_cache_query;
struct Query_cache_result;
class Query_cache;
/**
@brief This class represents a node in the linked chain of queries
belonging to one table.
@note The root of this linked list is not a query-type block, but the table-
type block which all queries has in common.
*/
struct Query_cache_block_table
{
Query_cache_block_table() {} /* Remove gcc warning */
TABLE_COUNTER_TYPE n; // numbr in table (from 0)
/**
This node holds a position in a static table list belonging
to the associated query (base 0).
*/
TABLE_COUNTER_TYPE n;
/**
Pointers to the next and previous node, linking all queries with
a common table.
*/
Query_cache_block_table *next, *prev;
/**
A pointer to the table-type block which all
linked queries has in common.
*/
Query_cache_table *parent;
/**
A method to calculate the address of the query cache block
owning this node. The purpose of this calculation is to
make it easier to move the query cache block without having
to modify all the pointer addresses.
*/
inline Query_cache_block *block();
};
struct Query_cache_block
{
Query_cache_block() {} /* Remove gcc warning */
......@@ -151,6 +178,11 @@ struct Query_cache_table
/* data need by some engines */
ulonglong engine_data_buff;
/**
The number of queries depending of this table.
*/
int32 m_cached_query_count;
inline char *db() { return (char *) data(); }
inline char *table() { return tbl; }
inline void table(char *table_arg) { tbl= table_arg; }
......@@ -237,9 +269,14 @@ class Query_cache
ulong free_memory, queries_in_cache, hits, inserts, refused,
free_memory_blocks, total_blocks, lowmem_prunes;
private:
pthread_cond_t COND_flush_finished;
bool flush_in_progress;
pthread_cond_t COND_cache_status_changed;
enum Cache_status { NO_FLUSH_IN_PROGRESS, FLUSH_IN_PROGRESS,
TABLE_FLUSH_IN_PROGRESS };
Cache_status m_cache_status;
void free_query_internal(Query_cache_block *point);
......@@ -253,7 +290,7 @@ class Query_cache
2. query block (for operation inside query (query block/results))
Thread doing cache flush releases the mutex once it sets
flush_in_progress flag, so other threads may bypass the cache as
m_cache_status flag, so other threads may bypass the cache as
if it is disabled, not waiting for reset to finish. The exception
is other threads that were going to do cache flush---they'll wait
till the end of a flush operation.
......@@ -270,6 +307,7 @@ class Query_cache
/* options */
ulong min_allocation_unit, min_result_data_size;
uint def_query_hash_size, def_table_hash_size;
uint mem_bin_num, mem_bin_steps; // See at init_cache & find_bin
my_bool initialized;
......@@ -295,10 +333,13 @@ class Query_cache
ulong data_len,
Query_cache_block *query_block,
my_bool first_block);
void invalidate_table(TABLE_LIST *table);
void invalidate_table(TABLE *table);
void invalidate_table(uchar *key, uint32 key_length);
void invalidate_table(Query_cache_block *table_block);
void invalidate_table(THD *thd, TABLE_LIST *table);
void invalidate_table(THD *thd, TABLE *table);
void invalidate_table(THD *thd, uchar *key, uint32 key_length);
void invalidate_table(THD *thd, Query_cache_block *table_block);
void invalidate_query_block_list(THD *thd,
Query_cache_block_table *list_root);
TABLE_COUNTER_TYPE
register_tables_from_list(TABLE_LIST *tables_used,
TABLE_COUNTER_TYPE counter,
......@@ -337,6 +378,8 @@ class Query_cache
Query_cache_block *pprev);
my_bool join_results(ulong join_limit);
void wait_while_table_flush_is_in_progress(bool *interrupt);
/*
Following function control structure_guard_mutex
by themself or don't need structure_guard_mutex
......@@ -347,8 +390,7 @@ class Query_cache
Query_cache_block *write_block_data(ulong data_len, uchar* data,
ulong header_len,
Query_cache_block::block_type type,
TABLE_COUNTER_TYPE ntab = 0,
my_bool under_guard=0);
TABLE_COUNTER_TYPE ntab = 0);
my_bool append_result_data(Query_cache_block **result,
ulong data_len, uchar* data,
Query_cache_block *parent);
......@@ -360,8 +402,7 @@ class Query_cache
inline ulong get_min_first_result_data_size();
inline ulong get_min_append_result_data_size();
Query_cache_block *allocate_block(ulong len, my_bool not_less,
ulong min,
my_bool under_guard=0);
ulong min);
/*
If query is cacheable return number tables in query
(query without tables not cached)
......@@ -424,6 +465,11 @@ class Query_cache
friend void query_cache_end_of_result(THD *thd);
friend void query_cache_abort(NET *net);
bool is_flushing(void)
{
return (m_cache_status != Query_cache::NO_FLUSH_IN_PROGRESS);
}
/*
The following functions are only used when debugging
We don't protect these with ifndef DBUG_OFF to not have to recompile
......
......@@ -1084,7 +1084,7 @@ static long mysql_rm_known_files(THD *thd, MY_DIR *dirp, const char *db,
}
}
if (thd->killed ||
(tot_list && mysql_rm_table_part2_with_lock(thd, tot_list, 1, 0, 1)))
(tot_list && mysql_rm_table_part2(thd, tot_list, 1, 0, 1, 1, 1)))
goto err;
/* Remove RAID directories */
......
......@@ -2445,7 +2445,7 @@ mysql_execute_command(THD *thd)
check_grant(thd, INSERT_ACL | CREATE_ACL, &new_list, 0, 1, 0)))
goto error;
}
query_cache_invalidate3(thd, first_table, 0);
if (end_active_trans(thd) || mysql_rename_tables(thd, first_table, 0))
goto error;
break;
......
......@@ -144,10 +144,14 @@ bool mysql_rename_tables(THD *thd, TABLE_LIST *table_list, bool silent)
}
}
VOID(pthread_mutex_lock(&LOCK_open));
if (lock_table_names(thd, table_list))
pthread_mutex_lock(&LOCK_open);
if (lock_table_names_exclusively(thd, table_list))
{
pthread_mutex_unlock(&LOCK_open);
goto err;
}
pthread_mutex_unlock(&LOCK_open);
error=0;
if ((ren_table=rename_tables(thd,table_list,0)))
{
......@@ -183,10 +187,14 @@ bool mysql_rename_tables(THD *thd, TABLE_LIST *table_list, bool silent)
send_ok(thd);
}
if (!error)
query_cache_invalidate3(thd, table_list, 0);
pthread_mutex_lock(&LOCK_open);
unlock_table_names(thd, table_list, (TABLE_LIST*) 0);
pthread_mutex_unlock(&LOCK_open);
err:
pthread_mutex_unlock(&LOCK_open);
/* enable logging back if needed */
if (disable_logs)
{
......
......@@ -1411,14 +1411,7 @@ bool mysql_rm_table(THD *thd,TABLE_LIST *tables, my_bool if_exists,
LOCK_open during wait_if_global_read_lock(), other threads could not
close their tables. This would make a pretty deadlock.
*/
thd->mysys_var->current_mutex= &LOCK_open;
thd->mysys_var->current_cond= &COND_refresh;
VOID(pthread_mutex_lock(&LOCK_open));
error= mysql_rm_table_part2(thd, tables, if_exists, drop_temporary, 0, 0);
pthread_mutex_unlock(&LOCK_open);
error= mysql_rm_table_part2(thd, tables, if_exists, drop_temporary, 0, 0, 1);
pthread_mutex_lock(&thd->mysys_var->mutex);
thd->mysys_var->current_mutex= 0;
thd->mysys_var->current_cond= 0;
......@@ -1433,49 +1426,6 @@ bool mysql_rm_table(THD *thd,TABLE_LIST *tables, my_bool if_exists,
DBUG_RETURN(FALSE);
}
/*
delete (drop) tables.
SYNOPSIS
mysql_rm_table_part2_with_lock()
thd Thread handle
tables List of tables to delete
if_exists If 1, don't give error if one table doesn't exists
dont_log_query Don't write query to log files. This will also not
generate warnings if the handler files doesn't exists
NOTES
Works like documented in mysql_rm_table(), but don't check
global_read_lock and don't send_ok packet to server.
RETURN
0 ok
1 error
*/
int mysql_rm_table_part2_with_lock(THD *thd,
TABLE_LIST *tables, bool if_exists,
bool drop_temporary, bool dont_log_query)
{
int error;
thd->mysys_var->current_mutex= &LOCK_open;
thd->mysys_var->current_cond= &COND_refresh;
VOID(pthread_mutex_lock(&LOCK_open));
error= mysql_rm_table_part2(thd, tables, if_exists, drop_temporary, 1,
dont_log_query);
pthread_mutex_unlock(&LOCK_open);
pthread_mutex_lock(&thd->mysys_var->mutex);
thd->mysys_var->current_mutex= 0;
thd->mysys_var->current_cond= 0;
pthread_mutex_unlock(&thd->mysys_var->mutex);
return error;
}
/*
Execute the drop of a normal or temporary table
......@@ -1508,7 +1458,7 @@ int mysql_rm_table_part2_with_lock(THD *thd,
int mysql_rm_table_part2(THD *thd, TABLE_LIST *tables, bool if_exists,
bool drop_temporary, bool drop_view,
bool dont_log_query)
bool dont_log_query, bool need_lock_open)
{
TABLE_LIST *table;
char path[FN_REFLEN], *alias;
......@@ -1520,9 +1470,11 @@ int mysql_rm_table_part2(THD *thd, TABLE_LIST *tables, bool if_exists,
String built_query;
DBUG_ENTER("mysql_rm_table_part2");
if (need_lock_open)
pthread_mutex_lock(&LOCK_open);
LINT_INIT(alias);
LINT_INIT(path_length);
safe_mutex_assert_owner(&LOCK_open);
if (thd->current_stmt_binlog_row_based && !dont_log_query)
{
......@@ -1555,8 +1507,15 @@ int mysql_rm_table_part2(THD *thd, TABLE_LIST *tables, bool if_exists,
}
}
if (!drop_temporary && lock_table_names(thd, tables))
if (!drop_temporary && lock_table_names_exclusively(thd, tables))
{
if (need_lock_open)
pthread_mutex_unlock(&LOCK_open);
DBUG_RETURN(1);
}
if (need_lock_open)
pthread_mutex_unlock(&LOCK_open);
/* Don't give warnings for not found errors, as we already generate notes */
thd->no_warnings_for_error= 1;
......@@ -1567,7 +1526,7 @@ int mysql_rm_table_part2(THD *thd, TABLE_LIST *tables, bool if_exists,
handlerton *table_type;
enum legacy_db_type frm_db_type;
mysql_ha_flush(thd, table, MYSQL_HA_CLOSE_FINAL, TRUE);
mysql_ha_flush(thd, table, MYSQL_HA_CLOSE_FINAL, !need_lock_open);
if (!close_temporary_table(thd, table))
{
tmp_table_deleted=1;
......@@ -1604,6 +1563,8 @@ int mysql_rm_table_part2(THD *thd, TABLE_LIST *tables, bool if_exists,
{
TABLE *locked_table;
abort_locked_tables(thd, db, table->table_name);
if (need_lock_open)
pthread_mutex_lock(&LOCK_open);
remove_table_from_cache(thd, db, table->table_name,
RTFC_WAIT_OTHER_THREAD_FLAG |
RTFC_CHECK_KILLED_FLAG);
......@@ -1614,6 +1575,9 @@ int mysql_rm_table_part2(THD *thd, TABLE_LIST *tables, bool if_exists,
if ((locked_table= drop_locked_tables(thd, db, table->table_name)))
table->table= locked_table;
if (need_lock_open)
pthread_mutex_unlock(&LOCK_open);
if (thd->killed)
{
thd->no_warnings_for_error= 0;
......@@ -1739,9 +1703,11 @@ int mysql_rm_table_part2(THD *thd, TABLE_LIST *tables, bool if_exists,
*/
}
}
if (!drop_temporary)
unlock_table_names(thd, tables, (TABLE_LIST*) 0);
if (need_lock_open)
pthread_mutex_lock(&LOCK_open);
unlock_table_names(thd, tables, (TABLE_LIST*) 0);
if (need_lock_open)
pthread_mutex_unlock(&LOCK_open);
thd->no_warnings_for_error= 0;
DBUG_RETURN(error);
}
......
......@@ -1268,8 +1268,6 @@ bool Table_triggers_list::drop_all_triggers(THD *thd, char *db, char *name)
bzero(&table, sizeof(table));
init_alloc_root(&table.mem_root, 8192, 0);
safe_mutex_assert_owner(&LOCK_open);
if (Table_triggers_list::check_n_load(thd, db, name, &table, 1))
{
result= 1;
......@@ -1431,26 +1429,24 @@ Table_triggers_list::change_table_name_in_trignames(const char *db_name,
}
/*
Update .TRG and .TRN files after renaming triggers' subject table.
/**
@brief Update .TRG and .TRN files after renaming triggers' subject table.
SYNOPSIS
change_table_name()
thd Thread context
db Old database of subject table
old_table Old name of subject table
new_db New database for subject table
new_table New name of subject table
@param[in,out] thd Thread context
@param[in] db Old database of subject table
@param[in] old_table Old name of subject table
@param[in] new_db New database for subject table
@param[in] new_table New name of subject table
NOTE
@note
This method tries to leave trigger related files in consistent state,
i.e. it either will complete successfully, or will fail leaving files
in their initial state.
Also this method assumes that subject table is not renamed to itself.
This method needs to be called under an exclusive table name lock.
RETURN VALUE
FALSE Success
TRUE Error
@retval FALSE Success
@retval TRUE Error
*/
bool Table_triggers_list::change_table_name(THD *thd, const char *db,
......@@ -1466,7 +1462,19 @@ bool Table_triggers_list::change_table_name(THD *thd, const char *db,
bzero(&table, sizeof(table));
init_alloc_root(&table.mem_root, 8192, 0);
safe_mutex_assert_owner(&LOCK_open);
uchar key[MAX_DBKEY_LENGTH];
uint key_length= (uint) (strmov(strmov((char*)&key[0], db)+1,
old_table)-(char*)&key[0])+1;
/*
This method interfaces the mysql server code protected by
either LOCK_open mutex or with an exclusive table name lock.
In the future, only an exclusive table name lock will be enough.
*/
#ifndef DBUG_OFF
if (!is_table_name_exclusively_locked_by_this_thread(thd, key, key_length))
safe_mutex_assert_owner(&LOCK_open);
#endif
DBUG_ASSERT(my_strcasecmp(table_alias_charset, db, new_db) ||
my_strcasecmp(table_alias_charset, old_table, new_table));
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment