Encoding fails with user created function

My hosting service runs MariaDB on a CentOS server. I connect via SSH (utf-8) or via PHP (with encoding everywhere set to utf-8 as well). The same issue happens no matter which of these approaches I use.

> show session variables like 'character_set%';

+--------------------------+------------------------------+

| Variable_name            | Value                        |

+--------------------------+------------------------------+

| character_set_client     | utf8                         |

| character_set_connection | utf8                         |

| character_set_database   | latin1                       |

| character_set_filesystem | binary                       |

| character_set_results    | utf8                         |

| character_set_server     | latin1                       |

| character_set_system     | utf8                         |

| character_sets_dir       | /usr/share/mariadb/charsets/ |

+--------------------------+------------------------------+

> DELIMITER //

> CREATE FUNCTION return_string( str TEXT ) RETURNS TEXT CHARSET 'utf8'

> BEGIN

> RETURN str;

> END //

> DELIMITER ;

> SELECT return_string('ā');

ERROR 1366 (22007): Incorrect string value: '\xC4\x81' for column ``.``.`str` at row 1

Unless I am mistaken, this is a bug. It works correctly in MySQL. MariaDB seems to get confused about which charset it is using with the user function. A simple "SELECT 'ā';" works correctly.

SHOW CREATE FUNCTION return_string;
shows the expected status of the function as being utf-8.

This works correctly if you use this instead:
CREATE FUNCTION return_string( str TEXT CHARSET 'utf8' ) RETURNS TEXT CHARSET 'utf8'
specifying the input encoding too. However, the connection was already in utf-8, the function is created to return values in utf-8 and the database tables (in the case where I discovered this issue) were in utf-8. So it seems really unexpected/buggy for it to try to interpret the function input as latin1. Especially when MySQL treats the input as utf-8 (at least in the slightly older version that I have access to).