Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash

Version for 5.0.
It fixes three problems:
1. The cause of the bug was that we did not check the table version for
 the HANDLER ... READ commands. We did not notice when a table was
 replaced by a new one. This can happen during ALTER TABLE, REPAIR
 TABLE, and OPTIMIZE TABLE (there might be more cases). I call the fix
 for this problem "the primary bug fix".
2. mysql_ha_flush() was not always called with a locked LOCK_open.
 Though the function comment clearly said it must.
 I changed the code so that the locking is done when required. I call
 the fix for this problem "the secondary fix".
3. In 5.0 (not in 4.1 or 4.0) DROP TABLE had a possible deadlock flaw in
 concur with FLUSH TABLES WITH READ LOCK. I call the fix for this
 problem "the 5.0 addendum fix".


include/my_pthread.h:
  Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
  Added a new macro for the 5.0 addendum fix.
mysql-test/r/handler.result:
  Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
  The test result.
mysql-test/t/handler.test:
  Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
  The test case.
sql/lock.cc:
  Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
  Changed a comment which did confuse me and which is not fully
  correct anymore after the 5.0 addendum fix.
  Added an assertion which would fire without the 5.0 addendum fix.
sql/mysql_priv.h:
  Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
  Changed a definition for the secondary fix.
sql/sql_base.cc:
  Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
  Changed function calls for the secondary fix.
sql/sql_class.cc:
  Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
  Changed a function call for the secondary fix.
sql/sql_handler.cc:
  Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
  The first two diffs make the primary bug fix.
  The rest is for the secondary fix.
sql/sql_table.cc:
  Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
  The first diff (four changed places) make the 5.0 addendum fix.
  The other three are changed function calls for the secondary fix.

include/my_pthread.h

@@ -536,9 +536,15 @@ void safe_mutex_end(FILE *file);
 #define pthread_cond_timedwait(A,B,C) safe_cond_timedwait((A),(B),(C),__FILE__,__LINE__)
 #define pthread_mutex_trylock(A) pthread_mutex_lock(A)
 #define pthread_mutex_t safe_mutex_t
-#define safe_mutex_assert_owner(mp) DBUG_ASSERT((mp)->count > 0 && pthread_equal(pthread_self(),(mp)->thread))
+#define safe_mutex_assert_owner(mp) \
+          DBUG_ASSERT((mp)->count > 0 && \
+                      pthread_equal(pthread_self(), (mp)->thread))
+#define safe_mutex_assert_not_owner(mp) \
+          DBUG_ASSERT(! (mp)->count || \
+                      ! pthread_equal(pthread_self(), (mp)->thread))
 #else
 #define safe_mutex_assert_owner(mp)
+#define safe_mutex_assert_not_owner(mp)
 #endif /* SAFE_MUTEX */
 
 	/* READ-WRITE thread locking */

mysql-test/r/handler.result

@@ -445,3 +445,40 @@ drop table t2;
 drop table t3;
 drop table t4;
 drop table t5;
+create table t1 (c1 int);
+insert into t1 values (1);
+handler t1 open;
+handler t1 read first;
+c1
+1
+send the below to another connection, do not wait for the result
+ optimize table t1;
+proceed with the normal connection
+handler t1 read next;
+c1
+1
+handler t1 close;
+read the result from the other connection
+Table	Op	Msg_type	Msg_text
+test.t1	optimize	status	OK
+proceed with the normal connection
+drop table t1;
+create table t1 (c1 int);
+insert into t1 values (14397);
+flush tables with read lock;
+drop table t1;
+ERROR HY000: Can't execute the query because you have a conflicting read lock
+send the below to another connection, do not wait for the result
+ drop table t1;
+proceed with the normal connection
+select * from t1;
+c1
+14397
+unlock tables;
+read the result from the other connection
+proceed with the normal connection
+select * from t1;
+ERROR 42S02: Table 'test.t1' doesn't exist
+drop table if exists t1;
+Warnings:
+Note	1051	Unknown table 't1'

mysql-test/t/handler.test

@@ -347,4 +347,79 @@ drop table t3;
 drop table t4;
 drop table t5;
 
+#
+# Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
+#
+create table t1 (c1 int);
+insert into t1 values (1);
+# client 1
+handler t1 open;
+handler t1 read first;
+# client 2
+connect (con2,localhost,root,,);
+connection con2;
+--exec echo send the below to another connection, do not wait for the result
+send optimize table t1;
+--sleep 1
+# client 1
+--exec echo proceed with the normal connection
+connection default;
+handler t1 read next;
+handler t1 close;
+# client 2
+--exec echo read the result from the other connection
+connection con2;
+reap;
+# client 1
+--exec echo proceed with the normal connection
+connection default;
+drop table t1;
+
 # End of 4.1 tests
+
+#
+# Addendum to Bug#14397 - OPTIMIZE TABLE with an open HANDLER causes a crash
+# Show that DROP TABLE can no longer deadlock against
+# FLUSH TABLES WITH READ LOCK. This is a 5.0 issue.
+#
+create table t1 (c1 int);
+insert into t1 values (14397);
+flush tables with read lock;
+# The thread with the global read lock cannot drop the table itself:
+--error 1223
+drop table t1;
+#
+# client 2
+# We need a second connection to try the drop.
+# The drop waits for the global read lock to go away.
+# Without the addendum fix it locked LOCK_open before entering the wait loop.
+connection con2;
+--exec echo send the below to another connection, do not wait for the result
+send drop table t1;
+--sleep 1
+#
+# client 1
+# Now we need something that wants LOCK_open. A simple table access which
+# opens the table does the trick.
+--exec echo proceed with the normal connection
+connection default;
+# This would hang on LOCK_open without the 5.0 addendum fix.
+select * from t1;
+# Release the read lock. This should make the DROP go through.
+unlock tables;
+#
+# client 2
+# Read the result of the drop command.
+connection con2;
+--exec echo read the result from the other connection
+reap;
+#
+# client 1
+# Now back to normal operation. The table should not exist any more.
+--exec echo proceed with the normal connection
+connection default;
+--error 1146
+select * from t1;
+# Just to be sure and not confuse the next test case writer.
+drop table if exists t1;
+

sql/lock.cc

@@ -815,10 +815,13 @@ static void print_lock_error(int error, const char *table)
 
   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 - see mysql_rm_table. Other mutexes could
-  be a problem too - grep the code for global_read_lock if you want to use
-  any other mutex here)
+  (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
@@ -937,6 +940,13 @@ bool wait_if_global_read_lock(THD *thd, bool abort_on_refresh,
   DBUG_ENTER("wait_if_global_read_lock");
 
   LINT_INIT(old_message);
+  /*
+    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.
+  */
+  safe_mutex_assert_not_owner(&LOCK_open);
+
   (void) pthread_mutex_lock(&LOCK_global_read_lock);
   if ((need_exit_cond= must_wait))
   {

sql/mysql_priv.h

@@ -890,7 +890,8 @@ bool mysql_ha_open(THD *thd, TABLE_LIST *tables, bool reopen);
 bool mysql_ha_close(THD *thd, TABLE_LIST *tables);
 bool mysql_ha_read(THD *, TABLE_LIST *,enum enum_ha_read_modes,char *,
                    List<Item> *,enum ha_rkey_function,Item *,ha_rows,ha_rows);
-int mysql_ha_flush(THD *thd, TABLE_LIST *tables, uint mode_flags);
+int mysql_ha_flush(THD *thd, TABLE_LIST *tables, uint mode_flags,
+                   bool is_locked);
 /* mysql_ha_flush mode_flags bits */
 #define MYSQL_HA_CLOSE_FINAL        0x00
 #define MYSQL_HA_REOPEN_ON_USAGE    0x01

sql/sql_base.cc

@@ -311,7 +311,8 @@ bool close_cached_tables(THD *thd, bool if_wait_for_refresh,
     thd->proc_info="Flushing tables";
 
     close_old_data_files(thd,thd->open_tables,1,1);
-    mysql_ha_flush(thd, tables, MYSQL_HA_REOPEN_ON_USAGE | MYSQL_HA_FLUSH_ALL);
+    mysql_ha_flush(thd, tables, MYSQL_HA_REOPEN_ON_USAGE | MYSQL_HA_FLUSH_ALL,
+                   TRUE);
     bool found=1;
     /* Wait until all threads has closed all the tables we had locked */
     DBUG_PRINT("info",
@@ -1238,7 +1239,7 @@ TABLE *open_table(THD *thd, TABLE_LIST *table_list, MEM_ROOT *mem_root,
 
   /* close handler tables which are marked for flush */
   if (thd->handler_tables)
-    mysql_ha_flush(thd, (TABLE_LIST*) NULL, MYSQL_HA_REOPEN_ON_USAGE);
+    mysql_ha_flush(thd, (TABLE_LIST*) NULL, MYSQL_HA_REOPEN_ON_USAGE, TRUE);
 
   for (table=(TABLE*) hash_search(&open_cache,(byte*) key,key_length) ;
        table && table->in_use ;
@@ -1642,7 +1643,7 @@ bool wait_for_tables(THD *thd)
   {
     thd->some_tables_deleted=0;
     close_old_data_files(thd,thd->open_tables,0,dropping_tables != 0);
-    mysql_ha_flush(thd, (TABLE_LIST*) NULL, MYSQL_HA_REOPEN_ON_USAGE);
+    mysql_ha_flush(thd, (TABLE_LIST*) NULL, MYSQL_HA_REOPEN_ON_USAGE, TRUE);
     if (!table_is_used(thd->open_tables,1))
       break;
     (void) pthread_cond_wait(&COND_refresh,&LOCK_open);

sql/sql_class.cc

@@ -376,7 +376,7 @@ void THD::cleanup(void)
     close_thread_tables(this);
   }
   mysql_ha_flush(this, (TABLE_LIST*) 0,
-                 MYSQL_HA_CLOSE_FINAL | MYSQL_HA_FLUSH_ALL);
+                 MYSQL_HA_CLOSE_FINAL | MYSQL_HA_FLUSH_ALL, FALSE);
   hash_free(&handler_tables_hash);
   delete_dynamic(&user_var_events);
   hash_free(&user_vars);

sql/sql_handler.cc

@@ -336,6 +336,7 @@ bool mysql_ha_read(THD *thd, TABLE_LIST *tables,
                    ha_rows select_limit_cnt, ha_rows offset_limit_cnt)
 {
   TABLE_LIST    *hash_tables;
+  TABLE         **table_ptr;
   TABLE         *table;
   MYSQL_LOCK    *lock;
   List<Item>	list;
@@ -368,6 +369,28 @@ bool mysql_ha_read(THD *thd, TABLE_LIST *tables,
     DBUG_PRINT("info-in-hash",("'%s'.'%s' as '%s' tab %p",
                                hash_tables->db, hash_tables->table_name,
                                hash_tables->alias, table));
+    /* Table might have been flushed. */
+    if (table && (table->s->version != refresh_version))
+    {
+      /*
+        We must follow the thd->handler_tables chain, as we need the
+        address of the 'next' pointer referencing this table
+        for close_thread_table().
+      */
+      for (table_ptr= &(thd->handler_tables);
+           *table_ptr && (*table_ptr != table);
+           table_ptr= &(*table_ptr)->next)
+      {}
+      (*table_ptr)->file->ha_index_or_rnd_end();
+      VOID(pthread_mutex_lock(&LOCK_open));
+      if (close_thread_table(thd, table_ptr))
+      {
+        /* Tell threads waiting for refresh that something has happened */
+        VOID(pthread_cond_broadcast(&COND_refresh));
+      }
+      VOID(pthread_mutex_unlock(&LOCK_open));
+      table= hash_tables->table= NULL;
+    }
     if (!table)
     {
       /*
@@ -594,6 +617,7 @@ err0:
                                 MYSQL_HA_REOPEN_ON_USAGE mark for reopen.
                                 MYSQL_HA_FLUSH_ALL flush all tables, not only
                                 those marked for flush.
+    is_locked                   If LOCK_open is locked.
 
   DESCRIPTION
     The list of HANDLER tables may be NULL, in which case all HANDLER
@@ -601,7 +625,6 @@ err0:
     If 'tables' is NULL and MYSQL_HA_FLUSH_ALL is not set,
     all HANDLER tables marked for flush are closed.
     Broadcasts a COND_refresh condition, for every table closed.
-    The caller must lock LOCK_open.
 
   NOTE
     Since mysql_ha_flush() is called when the base table has to be closed,
@@ -611,10 +634,12 @@ err0:
     0  ok
 */
 
-int mysql_ha_flush(THD *thd, TABLE_LIST *tables, uint mode_flags)
+int mysql_ha_flush(THD *thd, TABLE_LIST *tables, uint mode_flags,
+                   bool is_locked)
 {
   TABLE_LIST    *tmp_tables;
   TABLE         **table_ptr;
+  bool          did_lock= FALSE;
   DBUG_ENTER("mysql_ha_flush");
   DBUG_PRINT("enter", ("tables: %p  mode_flags: 0x%02x", tables, mode_flags));
 
@@ -640,6 +665,12 @@ int mysql_ha_flush(THD *thd, TABLE_LIST *tables, uint mode_flags)
                              (*table_ptr)->s->db,
                              (*table_ptr)->s->table_name,
                              (*table_ptr)->alias));
+          /* The first time it is required, lock for close_thread_table(). */
+          if (! did_lock && ! is_locked)
+          {
+            VOID(pthread_mutex_lock(&LOCK_open));
+            did_lock= TRUE;
+          }
           mysql_ha_flush_table(thd, table_ptr, mode_flags);
           continue;
         }
@@ -658,6 +689,12 @@ int mysql_ha_flush(THD *thd, TABLE_LIST *tables, uint mode_flags)
       if ((mode_flags & MYSQL_HA_FLUSH_ALL) ||
           ((*table_ptr)->s->version != refresh_version))
       {
+        /* The first time it is required, lock for close_thread_table(). */
+        if (! did_lock && ! is_locked)
+        {
+          VOID(pthread_mutex_lock(&LOCK_open));
+          did_lock= TRUE;
+        }
         mysql_ha_flush_table(thd, table_ptr, mode_flags);
         continue;
       }
@@ -665,6 +702,10 @@ int mysql_ha_flush(THD *thd, TABLE_LIST *tables, uint mode_flags)
     }
   }
 
+  /* Release the lock if it was taken by this function. */
+  if (did_lock)
+    VOID(pthread_mutex_unlock(&LOCK_open));
+
   DBUG_RETURN(0);
 }
 
@@ -696,8 +737,8 @@ static int mysql_ha_flush_table(THD *thd, TABLE **table_ptr, uint mode_flags)
                       table->alias, mode_flags));
 
   if ((hash_tables= (TABLE_LIST*) hash_search(&thd->handler_tables_hash,
-                                        (byte*) (*table_ptr)->alias,
-                                        strlen((*table_ptr)->alias) + 1)))
+                                        (byte*) table->alias,
+                                        strlen(table->alias) + 1)))
   {
     if (! (mode_flags & MYSQL_HA_REOPEN_ON_USAGE))
     {
@@ -712,6 +753,7 @@ static int mysql_ha_flush_table(THD *thd, TABLE **table_ptr, uint mode_flags)
   }    
 
   (*table_ptr)->file->ha_index_or_rnd_end();
+  safe_mutex_assert_owner(&LOCK_open);
   if (close_thread_table(thd, table_ptr))
   {
     /* Tell threads waiting for refresh that something has happened */

sql/sql_table.cc

@@ -103,23 +103,28 @@ bool mysql_rm_table(THD *thd,TABLE_LIST *tables, my_bool if_exists,
 
   /* mark for close and remove all cached entries */
 
-  thd->mysys_var->current_mutex= &LOCK_open;
-  thd->mysys_var->current_cond= &COND_refresh;
-  VOID(pthread_mutex_lock(&LOCK_open));
-
   if (!drop_temporary)
   {
     if ((error= wait_if_global_read_lock(thd, 0, 1)))
     {
       my_error(ER_TABLE_NOT_LOCKED_FOR_WRITE, MYF(0), tables->table_name);
-      goto err;
+      DBUG_RETURN(TRUE);
     }
     else
       need_start_waiters= TRUE;
   }
+
+  /*
+    Acquire LOCK_open after wait_if_global_read_lock(). If we would hold
+    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);
 
-err:
   pthread_mutex_unlock(&LOCK_open);
 
   pthread_mutex_lock(&thd->mysys_var->mutex);
@@ -232,7 +237,7 @@ int mysql_rm_table_part2(THD *thd, TABLE_LIST *tables, bool if_exists,
     char *db=table->db;
     db_type table_type= DB_TYPE_UNKNOWN;
 
-    mysql_ha_flush(thd, table, MYSQL_HA_CLOSE_FINAL);
+    mysql_ha_flush(thd, table, MYSQL_HA_CLOSE_FINAL, TRUE);
     if (!close_temporary_table(thd, db, table->table_name))
     {
       tmp_table_deleted=1;
@@ -2171,7 +2176,7 @@ static bool mysql_admin_table(THD* thd, TABLE_LIST* tables,
                             Protocol::SEND_NUM_ROWS | Protocol::SEND_EOF))
     DBUG_RETURN(TRUE);
 
-  mysql_ha_flush(thd, tables, MYSQL_HA_CLOSE_FINAL);
+  mysql_ha_flush(thd, tables, MYSQL_HA_CLOSE_FINAL, FALSE);
   for (table= tables; table; table= table->next_local)
   {
     char table_name[NAME_LEN*2+2];
@@ -3128,7 +3133,7 @@ bool mysql_alter_table(THD *thd,char *new_db, char *new_name,
     new_db= db;
   used_fields=create_info->used_fields;
 
-  mysql_ha_flush(thd, table_list, MYSQL_HA_CLOSE_FINAL);
+  mysql_ha_flush(thd, table_list, MYSQL_HA_CLOSE_FINAL, FALSE);
   /* DISCARD/IMPORT TABLESPACE is always alone in an ALTER TABLE */
   if (alter_info->tablespace_op != NO_TABLESPACE_OP)
     DBUG_RETURN(mysql_discard_or_import_tablespace(thd,table_list,