Merge chilla.local:/home/mydev/mysql-5.0

into  chilla.local:/home/mydev/mysql-5.0-bug16218

6 files changed, 91 insertions, 37 deletions

diff --git a/sql/field.cc b/sql/field.cc
index 7c25e4ad9f7..946351efe36 100644
--- a/sql/field.cc
+++ b/sql/field.cc
@@ -1515,7 +1515,8 @@ bool Field::optimize_range(uint idx, uint part)
 }
 
 
-Field *Field::new_field(MEM_ROOT *root, struct st_table *new_table)
+Field *Field::new_field(MEM_ROOT *root, struct st_table *new_table,
+                        bool keep_type __attribute__((unused)))
 {
   Field *tmp;
   if (!(tmp= (Field*) memdup_root(root,(char*) this,size_of())))
@@ -1540,7 +1541,7 @@ Field *Field::new_key_field(MEM_ROOT *root, struct st_table *new_table,
                             uint new_null_bit)
 {
   Field *tmp;
-  if ((tmp= new_field(root, new_table)))
+  if ((tmp= new_field(root, new_table, table == new_table)))
   {
     tmp->ptr=      new_ptr;
     tmp->null_ptr= new_null_ptr;
@@ -6227,29 +6228,21 @@ uint Field_string::max_packed_col_length(uint max_length)
 }
 
 
-Field *Field_string::new_field(MEM_ROOT *root, struct st_table *new_table)
+Field *Field_string::new_field(MEM_ROOT *root, struct st_table *new_table,
+                               bool keep_type)
 {
   Field *new_field;
 
-  if (type() != MYSQL_TYPE_VAR_STRING || table == new_table)
-    return Field::new_field(root, new_table);
+  if (type() != MYSQL_TYPE_VAR_STRING || keep_type)
+    return Field::new_field(root, new_table, keep_type);
 
   /*
     Old VARCHAR field which should be modified to a VARCHAR on copy
     This is done to ensure that ALTER TABLE will convert old VARCHAR fields
     to now VARCHAR fields.
   */
-  if ((new_field= new Field_varstring(field_length, maybe_null(),
-                                      field_name, new_table, charset())))
-  {
-    /*
-      delayed_insert::get_local_table() needs a ptr copied from old table.
-      This is what other new_field() methods do too. The above method of
-      Field_varstring sets ptr to NULL.
-    */
-    new_field->ptr= ptr;
-  }
-  return new_field;
+  return new Field_varstring(field_length, maybe_null(),
+                             field_name, new_table, charset());
 }
 
 /****************************************************************************
@@ -6741,9 +6734,11 @@ int Field_varstring::cmp_binary(const char *a_ptr, const char *b_ptr,
 }
 
 
-Field *Field_varstring::new_field(MEM_ROOT *root, struct st_table *new_table)
+Field *Field_varstring::new_field(MEM_ROOT *root, struct st_table *new_table,
+                                  bool keep_type)
 {
-  Field_varstring *res= (Field_varstring*) Field::new_field(root, new_table);
+  Field_varstring *res= (Field_varstring*) Field::new_field(root, new_table,
+                                                            keep_type);
   if (res)
     res->length_bytes= length_bytes;
   return res;
diff --git a/sql/field.h b/sql/field.h
index ed13372df71..3b33d3651e3 100644
--- a/sql/field.h
+++ b/sql/field.h
@@ -211,7 +211,8 @@ public:
   */
   virtual bool can_be_compared_as_longlong() const { return FALSE; }
   virtual void free() {}
-  virtual Field *new_field(MEM_ROOT *root, struct st_table *new_table);
+  virtual Field *new_field(MEM_ROOT *root, struct st_table *new_table,
+                           bool keep_type);
   virtual Field *new_key_field(MEM_ROOT *root, struct st_table *new_table,
                                char *new_ptr, uchar *new_null_ptr,
                                uint new_null_bit);
@@ -1033,7 +1034,7 @@ public:
   enum_field_types real_type() const { return FIELD_TYPE_STRING; }
   bool has_charset(void) const
   { return charset() == &my_charset_bin ? FALSE : TRUE; }
-  Field *new_field(MEM_ROOT *root, struct st_table *new_table);
+  Field *new_field(MEM_ROOT *root, struct st_table *new_table, bool keep_type);
 };
 
 
@@ -1105,7 +1106,7 @@ public:
   enum_field_types real_type() const { return MYSQL_TYPE_VARCHAR; }
   bool has_charset(void) const
   { return charset() == &my_charset_bin ? FALSE : TRUE; }
-  Field *new_field(MEM_ROOT *root, struct st_table *new_table);
+  Field *new_field(MEM_ROOT *root, struct st_table *new_table, bool keep_type);
   Field *new_key_field(MEM_ROOT *root, struct st_table *new_table,
                        char *new_ptr, uchar *new_null_ptr,
                        uint new_null_bit);
diff --git a/sql/sql_insert.cc b/sql/sql_insert.cc
index ba0d2d00f2c..5db387d557c 100644
--- a/sql/sql_insert.cc
+++ b/sql/sql_insert.cc
@@ -17,6 +17,44 @@
 
 /* Insert of records */
 
+/*
+  INSERT DELAYED
+
+  Insert delayed is distinguished from a normal insert by lock_type ==
+  TL_WRITE_DELAYED instead of TL_WRITE. It first tries to open a
+  "delayed" table (delayed_get_table()), but falls back to
+  open_and_lock_tables() on error and proceeds as normal insert then.
+
+  Opening a "delayed" table means to find a delayed insert thread that
+  has the table open already. If this fails, a new thread is created and
+  waited for to open and lock the table.
+
+  If accessing the thread succeeded, in
+  delayed_insert::get_local_table() the table of the thread is copied
+  for local use. A copy is required because the normal insert logic
+  works on a target table, but the other threads table object must not
+  be used. The insert logic uses the record buffer to create a record.
+  And the delayed insert thread uses the record buffer to pass the
+  record to the table handler. So there must be different objects. Also
+  the copied table is not included in the lock, so that the statement
+  can proceed even if the real table cannot be accessed at this moment.
+
+  Copying a table object is not a trivial operation. Besides the TABLE
+  object there are the field pointer array, the field objects and the
+  record buffer. After copying the field objects, their pointers into
+  the record must be "moved" to point to the new record buffer.
+
+  After this setup the normal insert logic is used. Only that for
+  delayed inserts write_delayed() is called instead of write_record().
+  It inserts the rows into a queue and signals the delayed insert thread
+  instead of writing directly to the table.
+
+  The delayed insert thread awakes from the signal. It locks the table,
+  inserts the rows from the queue, unlocks the table, and waits for the
+  next signal. It does normally live until a FLUSH TABLES or SHUTDOWN.
+
+*/
+
 #include "mysql_priv.h"
 #include "sp_head.h"
 #include "sql_trigger.h"
@@ -1441,6 +1479,7 @@ TABLE *delayed_insert::get_local_table(THD* client_thd)
   my_ptrdiff_t adjust_ptrs;
   Field **field,**org_field, *found_next_number_field;
   TABLE *copy;
+  DBUG_ENTER("delayed_insert::get_local_table");
 
   /* First request insert thread to get a lock */
   status=1;
@@ -1464,31 +1503,47 @@ TABLE *delayed_insert::get_local_table(THD* client_thd)
     }
   }
 
+  /*
+    Allocate memory for the TABLE object, the field pointers array, and
+    one record buffer of reclength size. Normally a table has three
+    record buffers of rec_buff_length size, which includes alignment
+    bytes. Since the table copy is used for creating one record only,
+    the other record buffers and alignment are unnecessary.
+  */
   client_thd->proc_info="allocating local table";
   copy= (TABLE*) client_thd->alloc(sizeof(*copy)+
 				   (table->s->fields+1)*sizeof(Field**)+
 				   table->s->reclength);
   if (!copy)
     goto error;
+
+  /* Copy the TABLE object. */
   *copy= *table;
   copy->s= &copy->share_not_to_be_used;
   // No name hashing
   bzero((char*) &copy->s->name_hash,sizeof(copy->s->name_hash));
   /* We don't need to change the file handler here */
 
-  field=copy->field=(Field**) (copy+1);
-  copy->record[0]=(byte*) (field+table->s->fields+1);
-  memcpy((char*) copy->record[0],(char*) table->record[0],table->s->reclength);
+  /* Assign the pointers for the field pointers array and the record. */
+  field= copy->field= (Field**) (copy + 1);
+  copy->record[0]= (byte*) (field + table->s->fields + 1);
+  memcpy((char*) copy->record[0], (char*) table->record[0],
+         table->s->reclength);
 
-  /* Make a copy of all fields */
-
-  adjust_ptrs=PTR_BYTE_DIFF(copy->record[0],table->record[0]);
+  /*
+    Make a copy of all fields.
+    The copied fields need to point into the copied record. This is done
+    by copying the field objects with their old pointer values and then
+    "move" the pointers by the distance between the original and copied
+    records. That way we preserve the relative positions in the records.
+  */
+  adjust_ptrs= PTR_BYTE_DIFF(copy->record[0], table->record[0]);
 
-  found_next_number_field=table->found_next_number_field;
-  for (org_field=table->field ; *org_field ; org_field++,field++)
+  found_next_number_field= table->found_next_number_field;
+  for (org_field= table->field; *org_field; org_field++, field++)
   {
-    if (!(*field= (*org_field)->new_field(client_thd->mem_root,copy)))
-      return 0;
+    if (!(*field= (*org_field)->new_field(client_thd->mem_root, copy, 1)))
+      DBUG_RETURN(0);
     (*field)->orig_table= copy;			// Remove connection
     (*field)->move_field(adjust_ptrs);		// Point at copy->record[0]
     if (*org_field == found_next_number_field)
@@ -1515,14 +1570,14 @@ TABLE *delayed_insert::get_local_table(THD* client_thd)
   /* Adjust lock_count. This table object is not part of a lock. */
   copy->lock_count= 0;
 
-  return copy;
+  DBUG_RETURN(copy);
 
   /* Got fatal error */
  error:
   tables_in_use--;
   status=1;
   pthread_cond_signal(&cond);			// Inform thread about abort
-  return 0;
+  DBUG_RETURN(0);
 }
 
 
diff --git a/sql/sql_select.cc b/sql/sql_select.cc
index 41ff9185cfb..400342854a6 100644
--- a/sql/sql_select.cc
+++ b/sql/sql_select.cc
@@ -8017,7 +8017,8 @@ Field* create_tmp_field_from_field(THD *thd, Field* org_field,
                                    org_field->field_name, table,
                                    org_field->charset());
   else
-    new_field= org_field->new_field(thd->mem_root, table);
+    new_field= org_field->new_field(thd->mem_root, table,
+                                    table == org_field->table);
   if (new_field)
   {
     if (item)
@@ -13062,7 +13063,7 @@ setup_copy_fields(THD *thd, TMP_TABLE_PARAM *param,
 	   saved value
 	*/
 	Field *field= item->field;
-	item->result_field=field->new_field(thd->mem_root,field->table);
+	item->result_field=field->new_field(thd->mem_root,field->table, 1);
 	char *tmp=(char*) sql_alloc(field->pack_length()+1);
 	if (!tmp)
 	  goto err;
diff --git a/sql/sql_trigger.cc b/sql/sql_trigger.cc
index 28d7dc0bb9d..2773dcb14ff 100644
--- a/sql/sql_trigger.cc
+++ b/sql/sql_trigger.cc
@@ -747,7 +747,8 @@ bool Table_triggers_list::prepare_record1_accessors(TABLE *table)
       QQ: it is supposed that it is ok to use this function for field
       cloning...
     */
-    if (!(*old_fld= (*fld)->new_field(&table->mem_root, table)))
+    if (!(*old_fld= (*fld)->new_field(&table->mem_root, table,
+                                      table == (*fld)->table)))
       return 1;
     (*old_fld)->move_field((my_ptrdiff_t)(table->record[1] -
                                           table->record[0]));
diff --git a/sql/table.cc b/sql/table.cc
index 83711577699..728b98b2d35 100644
--- a/sql/table.cc
+++ b/sql/table.cc
@@ -804,7 +804,8 @@ int openfrm(THD *thd, const char *name, const char *alias, uint db_stat,
 	    if (!(field->flags & BLOB_FLAG))
 	    {					// Create a new field
 	      field=key_part->field=field->new_field(&outparam->mem_root,
-						     outparam);
+                                                     outparam,
+                                                     outparam == field->table);
 	      field->field_length=key_part->length;
 	    }
 	  }