BUG#11763712 - 56458: KILLING A FLUSH TABLE FOR A MERGE/CHILD

                      CRASHES SERVER

Flushing of MERGE table or one of its child tables, which was
locked by flushing thread using LOCK TABLES, might have caused
crashes or assertion failures if the thread failed to reopen
child or parent table.
Particularly, this might have happened when another connection
killed this FLUSH TABLE statement/connection.
Also this problem might have occurred when we failed to reopen
MERGE table or one of its children when executing DDL statement
under LOCK TABLES.

The problem was caused by the fact that reopen_tables() might
have failed to reopen child table but still tried to reopen,
reattach children for and re-lock its parent. Vice versa it
might have failed to reopen parent but kept references from
children to parent around. Since reopen_tables() closes table
it has failed to reopen and therefore frees all associated
memory such dangling references led to crashes when followed.

This patch solves this problem by ensuring that we always close
parent table and all its children if we fail to reopen this
table or one of its children. Same happens if we fail to reattach
children to parent.

Affects 5.1 only.

1 files changed, 141 insertions, 42 deletions

diff --git a/sql/sql_base.cc b/sql/sql_base.cc
index dc78f3b84c6..971a54d88c2 100644
--- a/sql/sql_base.cc
+++ b/sql/sql_base.cc
@@ -96,6 +96,13 @@ static TABLE_SHARE *oldest_unused_share, end_of_unused_share;
 static pthread_mutex_t LOCK_table_share;
 static bool table_def_inited= 0;
 
+/**
+  Dummy TABLE instance which is used in reopen_tables() and reattach_merge()
+  functions to mark MERGE tables and their children with which there is some
+  kind of problem and which therefore we need to close.
+*/
+static TABLE bad_merge_marker;
+
 static int open_unireg_entry(THD *thd, TABLE *entry, TABLE_LIST *table_list,
 			     const char *alias,
                              char *cache_key, uint cache_key_length,
@@ -3215,46 +3222,65 @@ void close_data_files_and_morph_locks(THD *thd, const char *db,
 
 
 /**
+  @brief Mark merge parent and children with bad_merge_marker
+
+  @param[in,out]  parent       the TABLE object of the parent
+*/
+
+static void mark_merge_parent_and_children_as_bad(TABLE *parent)
+{
+  TABLE_LIST *child_l;
+  DBUG_ENTER("mark_merge_parent_and_children_as_bad");
+  parent->parent= &bad_merge_marker;
+  for (child_l= parent->child_l; ; child_l= child_l->next_global)
+  {
+    child_l->table->parent= &bad_merge_marker;
+    child_l->table= NULL;
+    if (&child_l->next_global == parent->child_last_l)
+      break;
+  }
+  DBUG_VOID_RETURN;
+}
+
+
+/**
   Reattach MERGE children after reopen.
 
   @param[in]     thd            thread context
-  @param[in,out] err_tables_p   pointer to pointer of tables in error
+
+  @note If reattach failed for certain MERGE table, the table (and all
+        it's children) are marked with bad_merge_marker.
 
   @return       status
-    @retval     FALSE           OK, err_tables_p unchanged
-    @retval     TRUE            Error, err_tables_p contains table(s)
+    @retval     FALSE           OK
+    @retval     TRUE            Error
 */
 
-static bool reattach_merge(THD *thd, TABLE **err_tables_p)
+static bool reattach_merge(THD *thd)
 {
   TABLE *table;
-  TABLE *next;
-  TABLE **prv_p= &thd->open_tables;
   bool error= FALSE;
   DBUG_ENTER("reattach_merge");
 
-  for (table= thd->open_tables; table; table= next)
+  for (table= thd->open_tables; table; table= table->next)
   {
-    next= table->next;
-    DBUG_PRINT("tcache", ("check table: '%s'.'%s' 0x%lx  next: 0x%lx",
+    DBUG_PRINT("tcache", ("check table: '%s'.'%s' 0x%lx",
                           table->s->db.str, table->s->table_name.str,
-                          (long) table, (long) next));
-    /* Reattach children for MERGE tables with "closed data files" only. */
-    if (table->child_l && !table->children_attached)
+                          (long) table));
+    /*
+      Reattach children only for MERGE tables that had children or parent
+      with "closed data files" and were reopen. For extra safety skip MERGE
+      tables which we failed to reopen (should not happen with current code).
+    */
+    if (table->child_l && table->parent != &bad_merge_marker &&
+        !table->children_attached)
     {
       DBUG_PRINT("tcache", ("MERGE parent, attach children"));
-      if(table->file->extra(HA_EXTRA_ATTACH_CHILDREN))
+      if (table->file->extra(HA_EXTRA_ATTACH_CHILDREN))
       {
         my_error(ER_CANT_REOPEN_TABLE, MYF(0), table->alias);
         error= TRUE;
-        /* Remove table from open_tables. */
-        *prv_p= next;
-        if (next)
-          prv_p= &next->next;
-        /* Stack table on error list. */
-        table->next= *err_tables_p;
-        *err_tables_p= table;
-        continue;
+        mark_merge_parent_and_children_as_bad(table);
       }
       else
       {
@@ -3264,7 +3290,6 @@ static bool reattach_merge(THD *thd, TABLE **err_tables_p)
                             table->s->table_name.str, (long) table));
       }
     }
-    prv_p= &table->next;
   }
   DBUG_RETURN(error);
 }
@@ -3294,7 +3319,6 @@ bool reopen_tables(THD *thd, bool get_locks, bool mark_share_as_old)
 {
   TABLE *table,*next,**prev;
   TABLE **tables,**tables_ptr;			// For locks
-  TABLE *err_tables= NULL;
   bool error=0, not_used;
   bool merge_table_found= FALSE;
   const uint flags= MYSQL_LOCK_NOTIFY_IF_NEED_REOPEN |
@@ -3328,29 +3352,69 @@ bool reopen_tables(THD *thd, bool get_locks, bool mark_share_as_old)
   for (table=thd->open_tables; table ; table=next)
   {
     uint db_stat=table->db_stat;
+    TABLE *parent= table->child_l ? table : table->parent;
     next=table->next;
     DBUG_PRINT("tcache", ("open table: '%s'.'%s' 0x%lx  "
                           "parent: 0x%lx  db_stat: %u",
                           table->s->db.str, table->s->table_name.str,
                           (long) table, (long) table->parent, db_stat));
-    if (table->child_l && !db_stat)
+    /*
+      If we need to reopen child or parent table in a MERGE table, then
+      children in this MERGE table has to be already detached at this
+      point.
+    */
+    DBUG_ASSERT(db_stat || !parent || !parent->children_attached);
+    /*
+      Thanks to the above assumption the below condition will guarantee that
+      merge_table_found is TRUE when we need to reopen child or parent table.
+      Note that it works even in situation when it is only a child and not a
+      parent that needs reopen (this can happen when get_locks == FALSE).
+    */
+    if (table->child_l && !table->children_attached)
       merge_table_found= TRUE;
-    if (!tables || (!db_stat && reopen_table(table)))
+
+    if (!tables)
     {
-      my_error(ER_CANT_REOPEN_TABLE, MYF(0), table->alias);
       /*
-        If we could not allocate 'tables', we may close open tables
-        here. If a MERGE table is affected, detach the children first.
-        It is not necessary to clear the child or parent table reference
-        of this table because the TABLE is freed. But we need to clear
-        the child or parent references of the other belonging tables so
-        that they cannot be moved into the unused_tables chain with
-        these pointers set.
+        If we could not allocate 'tables' we close ALL open tables here.
+        Before closing MERGE child or parent we need to detach children
+        and/or clear references in/to them.
       */
-      if (table->child_l || table->parent)
+      if (parent)
         detach_merge_children(table, TRUE);
-      VOID(hash_delete(&open_cache,(uchar*) table));
-      error=1;
+    }
+    else if (table->parent == &bad_merge_marker)
+    {
+      /*
+        This is either a child or a parent of a MERGE table for which
+        we already decided that we are unable to reopen it. Close it.
+
+        Reset parent reference, it may be used while freeing the table.
+      */
+      table->parent= NULL;
+    }
+    else if (!db_stat && reopen_table(table))
+    {
+      /*
+        If we fail to reopen a child or a parent in a MERGE table and the
+        MERGE table is affected for the first time, mark all relevant tables
+        invalid. Otherwise handle it as usual.
+
+        All in all we must end up with:
+        - child tables are detached from parent. This was done earlier,
+          but child<->parent references were kept valid for reopen.
+        - parent is not in the to-be-locked tables
+        - all child tables and parent are not in the THD::open_tables.
+        - all child tables and parent are not in the open_cache.
+
+        Please note that below we do additional pass through THD::open_tables
+        list to achieve the last three points.
+      */
+      if (parent)
+      {
+        mark_merge_parent_and_children_as_bad(parent);
+        table->parent= NULL;
+      }
     }
     else
     {
@@ -3366,21 +3430,56 @@ bool reopen_tables(THD *thd, bool get_locks, bool mark_share_as_old)
 	table->s->version=0;
 	table->open_placeholder= 0;
       }
+      continue;
     }
+    my_error(ER_CANT_REOPEN_TABLE, MYF(0), table->alias);
+    VOID(hash_delete(&open_cache, (uchar *) table));
+    error= 1;
   }
   *prev=0;
   /*
     When all tables are open again, we can re-attach MERGE children to
-    their parents. All TABLE objects are still present.
+    their parents.
+
+    If there was an error while reopening a child or a parent of a MERGE
+    table, or while reattaching child tables to their parents, some tables
+    may have been kept open but marked for close with bad_merge_marker.
+    Close these tables now.
   */
-  DBUG_PRINT("tcache", ("re-attaching MERGE tables: %d", merge_table_found));
-  if (!error && merge_table_found && reattach_merge(thd, &err_tables))
+  if (tables && merge_table_found && (error|= reattach_merge(thd)))
   {
-    while (err_tables)
+    prev= &thd->open_tables;
+    for (table= thd->open_tables; table; table= next)
     {
-      VOID(hash_delete(&open_cache, (uchar*) err_tables));
-      err_tables= err_tables->next;
+      next= table->next;
+      if (table->parent == &bad_merge_marker)
+      {
+        /* Remove merge parent from to-be-locked tables array. */
+        if (get_locks && table->child_l)
+        {
+          TABLE **t;
+          for (t= tables; t < tables_ptr; t++)
+          {
+            if (*t == table)
+            {
+              tables_ptr--;
+              memmove(t, t + 1, (tables_ptr - t) * sizeof(TABLE *));
+              break;
+            }
+          }
+        }
+        /* Reset parent reference, it may be used while freeing the table. */
+        table->parent= NULL;
+        /* Free table. */
+        VOID(hash_delete(&open_cache, (uchar *) table));
+      }
+      else
+      {
+        *prev= table;
+        prev= &table->next;
+      }
     }
+    *prev= 0;
   }
   DBUG_PRINT("tcache", ("open tables to lock: %u",
                         (uint) (tables_ptr - tables)));