The security model of SELinux describes all the access control rules as relationships between a subject entity (typically, a client of the database) and an object entity (such as a database object), each of which is identified by a security label. If access to an unlabeled object is attempted, the object is treated as if it were assigned the label unlabeled_t.

Currently, sepgsql allows security labels to be assigned to schemas, tables, columns, sequences, views, and functions. When sepgsql is in use, security labels are automatically assigned to supported database objects at creation time. This label is called a default security label, and is decided according to the system security policy, which takes as input the creator's label, the label assigned to the new object's parent object and optionally name of the constructed object.

A new database object basically inherits the security label of the parent object, except when the security policy has special rules known as type-transition rules, in which case a different label may be applied. For schemas, the parent object is the current database; for tables, sequences, views, and functions, it is the containing schema; for columns, it is the containing table.

For tables, db_table:select, db_table:insert, db_table:update or db_table:delete are checked for all the referenced target tables depending on the kind of statement; in addition, db_table:select is also checked for all the tables that contain columns referenced in the WHERE or RETURNING clause, as a data source for UPDATE, and so on.

Column-level permissions will also be checked for each referenced column. db_column:select is checked on not only the columns being read using SELECT, but those being referenced in other DML statements; db_column:update or db_column:insert will also be checked for columns being modified by UPDATE or INSERT.

For example, consider:

UPDATE t1 SET x = 2, y = md5sum(y) WHERE z = 100;

Here, db_column:update will be checked for t1.x, since it is being updated, db_column:{select update} will be checked for t1.y, since it is both updated and referenced, and db_column:select will be checked for t1.z, since it is only referenced. db_table:{select update} will also be checked at the table level.

For sequences, db_sequence:get_value is checked when we reference a sequence object using SELECT; however, note that we do not currently check permissions on execution of corresponding functions such as lastval().

For views, db_view:expand will be checked, then any other required permissions will be checked on the objects being expanded from the view, individually.

For functions, db_procedure:{execute} will be checked when user tries to execute a function as a part of query, or using fast-path invocation. If this function is a trusted procedure, it also checks db_procedure:{entrypoint} permission to check whether it can perform as entry point of trusted procedure.

In order to access any schema object, db_schema:search permission is required on the containing schema. When an object is referenced without schema qualification, schemas on which this permission is not present will not be searched (just as if the user did not have USAGE privilege on the schema). If an explicit schema qualification is present, an error will occur if the user does not have the requisite permission on the named schema.

The client must be allowed to access all referenced tables and columns, even if they originated from views which were then expanded, so that we apply consistent access control rules independent of the manner in which the table contents are referenced.

The default database privilege system allows database superusers to modify system catalogs using DML commands, and reference or modify toast tables. These operations are prohibited when sepgsql is enabled.

SELinux defines several permissions to control common operations for each object type; such as creation, alter, drop and relabel of security label. In addition, several object types have special permissions to control their characteristic operations; such as addition or deletion of name entries within a particular schema.

Creating a new database object requires create permission. SELinux will grant or deny this permission based on the client's security label and the proposed security label for the new object. In some cases, additional privileges are required:

When DROP command is executed, drop will be checked on the object being removed. Permissions will be also checked for objects dropped indirectly via CASCADE. Deletion of objects contained within a particular schema (tables, views, sequences and procedures) additionally requires remove_name on the schema.

When ALTER command is executed, setattr will be checked on the object being modified for each object types, except for subsidiary objects such as the indexes or triggers of a table, where permissions are instead checked on the parent object. In some cases, additional permissions are required:

Trusted procedures are similar to security definer functions or setuid commands. SELinux provides a feature to allow trusted code to run using a security label different from that of the client, generally for the purpose of providing highly controlled access to sensitive data (e.g., rows might be omitted, or the precision of stored values might be reduced). Whether or not a function acts as a trusted procedure is controlled by its security label and the operating system security policy. For example:

postgres=# CREATE TABLE customer (
               cid     int primary key,
               cname   text,
               credit  text
           );
CREATE TABLE
postgres=# SECURITY LABEL ON COLUMN customer.credit
               IS 'system_u:object_r:sepgsql_secret_table_t:s0';
SECURITY LABEL
postgres=# CREATE FUNCTION show_credit(int) RETURNS text
             AS 'SELECT regexp_replace(credit, ''-[0-9]+$'', ''-xxxx'', ''g'')
                        FROM customer WHERE cid = $1'
           LANGUAGE sql;
CREATE FUNCTION
postgres=# SECURITY LABEL ON FUNCTION show_credit(int)
               IS 'system_u:object_r:sepgsql_trusted_proc_exec_t:s0';
SECURITY LABEL

The above operations should be performed by an administrative user.

postgres=# SELECT * FROM customer;
ERROR:  SELinux: security policy violation
postgres=# SELECT cid, cname, show_credit(cid) FROM customer;
 cid | cname  |     show_credit
-----+--------+---------------------
   1 | taro   | 1111-2222-3333-xxxx
   2 | hanako | 5555-6666-7777-xxxx
(2 rows)

In this case, a regular user cannot reference customer.credit directly, but a trusted procedure show_credit allows him to print the credit card numbers of customers with some of the digits masked out.

It is possible to use SELinux's dynamic domain transition feature to switch the security label of the client process, the client domain, to a new context, if that is allowed by the security policy. The client domain needs the setcurrent permission and also dyntransition from the old to the new domain.

Dynamic domain transitions should be considered carefully, because they allow users to switch their label, and therefore their privileges, at their option, rather than (as in the case of a trusted procedure) as mandated by the system. Thus, the dyntransition permission is only considered safe when used to switch to a domain with a smaller set of privileges than the original one. For example:

regression=# select sepgsql_getcon();
                    sepgsql_getcon
-------------------------------------------------------
 unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023
(1 row)

regression=# SELECT sepgsql_setcon('unconfined_u:unconfined_r:unconfined_t:s0-s0:c1.c4');
 sepgsql_setcon 
----------------
 t
(1 row)

regression=# SELECT sepgsql_setcon('unconfined_u:unconfined_r:unconfined_t:s0-s0:c1.c1023');
ERROR:  SELinux: security policy violation

In this example above we were allowed to switch from the larger MCS range c1.c1023 to the smaller range c1.c4, but switching back was denied.

A combination of dynamic domain transition and trusted procedure enables an interesting use case that fits the typical process life-cycle of connection pooling software. Even if your connection pooling software is not allowed to run most of SQL commands, you can allow it to switch the security label of the client using the sepgsql_setcon() function from within a trusted procedure; that should take some credential to authorize the request to switch the client label. After that, this session will have the privileges of the target user, rather than the connection pooler. The connection pooler can later revert the security label change by again using sepgsql_setcon() with NULL argument, again invoked from within a trusted procedure with appropriate permissions checks. The point here is that only the trusted procedure actually has permission to change the effective security label, and only does so when given proper credentials. Of course, for secure operation, the credential store (table, procedure definition, or whatever) must be protected from unauthorized access.

We reject the LOAD command across the board, because any module loaded could easily circumvent security policy enforcement.