From PostgreSQL wiki
This page is for discussing the implementation of Row-Level Security (RLS) in PostgreSQL. You can find the current progress of this proposal at the commitfest page.
RLS is a security feature which allows users to give access to a sub-set of data in their table to others. Traditional RDBMS permission systems don't distinguish the individual rows in a table- GRANT SELECT on a table will allow a user to access all rows of that table. With RLS, requirements such as "a manager can only view sensetive employee information for those employees who report to them." can be specified and enforced by the database.
Jobs of access control feature
According to the definition of ISO/IEC27001 (information security management), design target of information security feature is to ensure confidentiality, integrity and availability of information asset. In short, these are often called C.I.A. Access control contributes towards both confidentiality and integrity. Access controls prevent unprivileged users from reading or writing information asset base on the rules which are configured in the access control system. Information or data itself does not have a particularly tangible form and therefore it must be stored in an object. Usually, access control features allow or prohibit users access to the object that contains the information or data. The intent of this feature (RLS) is to allow a more fine-grained control over the information inside of the objects. For example, regular GRANT/REVOKE mechanisms control access on the specified database object according to the access control list, but they do not allow anything more granular.
The purpose of row-level security feature is to prevent users (not means database roles, just users) from accessing unprivileged rows. Please note that "access" means two different direction of information (1) data read (rows => users) for confidentiality, and (2) data write (users => rows) for data integrity. Due to the nature of database access, it is the most straight-forwards way to describe the rule with regular qualifier style of WHERE clause; that is an expression returning a boolean value.
Overall, RLS prevents users from reading and writing rows that do not satisfy the rule (also know as a row-security policy). If we support per-command configuration, the row-security policy to be checked depends on the command. RLS design accepts individual row-security policy to be applied on SELECT, INSERT, UPDATE or DELETE. Relevant discussions are below.
We have some SQL commands that allow users to access database rows; SELECT, INSERT, UPDATE or DELETE. COPY TO/FROM is synonym of SELECT and INSERT from security perspective. In case of SELECT (data read), what we should do is quite simple: any rows that don't match with the configured row-security policy shall be filtered out for unprivileged users. In case of INSERT or UPDATE (data write), RLS prevents unprivileged rows from being written to the table, as if CHECK constraint performing. In case of UPDATE or DELETE (data write), RLS prevents unprivileged rows from appearing as candidates of modification; that means row-security policy should be applied on the table scan stage. We also need to pay attention on potential information leaks using the leaky-view scenario below. So, UPDATE and DELETE shall also take row-security checks of SELECT command on table scanning stage.
TRUNCATE command performs as if DELETE, but much faster. It has its own permission separated from DELETE. So, we re-define meaning of TRUNCATE permission; that also implies to ignore row-security policy of DELETE.
Row-security policy is set by table owner, using the following syntax. If no "ON <command>" given, a unique security policy shall be applied on all the supported privileges (SELECT, INSERT, UPDATE and DELETE).
For example, an UPDATE command requires SELECT privilege to read a row, yet UPDATE privilege to change a row.
ALTER TABLE <relname> SET ROW SECURITY FOR <privilege> TO (<expression>); ALTER TABLE <relname> RESET ROW SECURITY FOR <privilege>;
<privilege> can be one of: ALL, SELECT, INSERT, UPDATE, DELETE. The initial implementation may only support 'ALL'.
'ALL' would simply define the same policy for all the commands. Later commands which update only a single command would replace the policy for that command only, leaving the policy for the other commands to what was set from the initial 'ALL'. The same goes for a RESET. Listing the policies in a table form would look like:
ALTER TABLE mytable SET ROW SECURITY FOR ALL TO (where user = 1);
|mytable||select||user = 1|
|mytable||insert||user = 1|
|mytable||update||user = 1|
|mytable||delete||user = 1|
If it was followed by:
ALTER TABLE mytable SET ROW SECURITY FOR SELECT TO (where user = 2); the result would be:
|mytable||select||user = 2|
|mytable||insert||user = 1|
|mytable||update||user = 1|
|mytable||delete||user = 1|
and so on. If no row security policy has been set for a given command type (or the security policy has been reset for that command type), no row-level security would apply and regular GRANT permissions for the operation would be used. Note that row-level security does not override existing GRANT permissions but provide a more fine-grained control. For example, setting 'ROW SECURITY FOR SELECT' to allow a given user to give rows would only give that user access if the user ALSO has 'SELECT' privileges on the column or table in question.
Superuser bypass row-security policy (1) to avoid Trojan-horse attack by table owner (2) to ensure consistent view for database backup, however, row-security policy injected by extension (such as sepgsql) is exception.
Issues & discussion
Per-command security policy
Asymmetric row-security policy may cause unexpected information leaks using UPDATE or DELETE command with RETURNING clause or leaky functions in WHERE clause. It is an idea to append row-security policy of SELECT when executor scan the result relation. It ensures the rows to be modified are also visible to the user executing the UPDATE. One other idea was to enforce a unique policy for all commands, however, it has a problematic scenario when user wants to define INSERT-only relation.
Now we can implement writer-side checks on INSERT or UPDATE command using before-row triggers. On the other hand, it makes users to synchronize the configuration of RLS with triggers of this checks. One model case to solve this concern is implementation of FK constraints; that automatically defines before-row triggers that checks newer version of tuples to be inserted or updated.
pg_statistic hold some example of values on the table being analyzed. Right now, we have no way to prevent users to see collected values on statistical board. An idea is to mask the field if the relation has RLS policy.
Minimal core feature set
- Checks are only applied on table scanning. If writer-side checks are required, users can do that using triggers, even though it takes complex setting.
- A unique security policy can be configurable on a table. Even though RLS design allows per-command policy, we need to investigate whether asymmetric policy is harmless.
Previous discussion in 2010
- Before we can try to tackle row-level security generally, using labels or not, we need to fix the issues with data leaks in views.
- Related topics in -hackers
Issue: A leaky VIEWs for RLS
- This issue is summarized as: an untrusted user can define a function which stores all information it is presented, then query a view using that function in a way which will convince the planner to send every row of the underlying table to the function, thus leaking data in the table which the view was intended to prevent.
In this message, KaiGai pointed out we have two different causes of the problem, but both of them can cause same information leaks.
-  Unexpected order to evaluate qualifiers on a certain scan plan
- When a scan-plan has multiple qualifiers to filter scanned tuples, the optimizer sort the qualifiers based on their estimated cost to execute. If an exogenetic function has smaller cost than other qualifiers come from view definition, the exogenetic function shall be evaluated earlier than others, then contents of invisible tuples may be provided to malicious user-defined functions.
- It is reordered at order_qual_clauses().
-  Unexpected qualifier distributions into inside of join loop
- When planner makes a scan plan, it tries to distribute qualifiers of scans into smaller unit as possible as it can. For example, if a function takes arguments only come from a certain table, it will be distributed into scan plan of the table, not outside of the join.
- When a view contains JOIN clause between A and B, user can reference the view with his own WHERE clause. If a clause takes arguments depending on only A, the planner distribute the clause into the scan plan of A. In the reault, tuples to be filtered out may become visible to user defined functions.
- It is distributed at distribute_qual_to_rels().
- At the point , if we prevent all exogenetic functions to push down into join loops, it will make unignorable performance degradation, although the view might not be intended to security purpose.
- It needs a way to provide a hint whether the view is defined for security purpose, or not.
- Tom Lane suggested CREATE SECURITY VIEW AS ... statement.
- It was not concluded which is the default. Security view? Regular view?
- How about a GUC option to specify the default? NACKed.
- In addition, Robert Haas suggested to test privileges of users whether they have privileges to reference underlaying tables without vires, or not. If available, it is eventually harmless even if user defined functions are evaluated within join loop.
- Here was one opposition because this check will be applied on planner stage, not execution stage.
- KaiGai submitted a proof of concept patch that prevents to push down exogenetic functions into security views.
- At the point , we don't have any active discussions yet.
Discussions of RLS in PG
- Current thread on -hackers
- Josh Berkus on RLS in PG, Part 1
- Josh Berkus on RLS in PG, Part 2
- SEPostgreSQL_Specifications Specifications for SEPostgreSQL, includes RLS
Articles/Documentation of existing RLS implementations
- Oracle RLS Article, Part 1
- Oracle RLS Article, Part 2
- Oracle RLS and VPD Article
- SQL Server RLS with Classified Systems
- IBM/DB2 RLS Documentation
- PCI Compliant implementations
- Classified Environments
Components of an implementation
- Allow the query to be modified, prior to being passed to the planner, in such a way that the rows returned will be those the user is authorized for
- This depends on being able to tell the planner that this filter must be done, in some way, prior to user-defined functions being called
- This issue is related to the VIEW leak described above. Once that issue has been resolved, this should be pretty straight-forward to implement
- Grammar for catalog updates/changes; user-interface for specifying how RLS is to be done
- Catalog changes for storing RLS information
- Storage - Could this just be a regular column in the table? It would be good to avoid changing the header or creating a system column for this.
- We would need to track, in some fashion, the "security" column in the catalog, perhaps as a flag on pg_attribute, or a 'security_attnum' in pg_class, etc.
- Planner updates to enforce the filter based on RLS- this can't be done till after we deal with the issue with VIEWs
- Executor changes may not be required.. but how to deal with stored plans? Use invalidation if anything changes with regard to RLS?
- Covert Channel
- If we try to insert a value which violates a PK constraint, we can assume existence of invisible PK from the error.
- The same issue exists in a Foreign Key relationship situation
- Other databases with row-level security don't address this issue, so it's unclear if we really need to (reference?)
- In any case, this isn't something we need to address in our initial RLS implementation
- Order to evaluate row level policy
- Addressed above with regard to views, once we solve that, this will be handled
- TRUNCATE statement
- TRUNCATE is expected to be fast.
- If the user does not have rights to remove all rows from the table regardless of row-level policy, then any TRUNCATE must be denied.
- Turning TRUNCATE (a PostgreSQL extension which is not in the SQL spec anyway) into a DELETE FROM doesn't make sense.
- COPY TO statement
- COPY can just be reformed into a COPY statement with a query being used instead, eg:
- COPY tblname TO xxx; can be replaced by COPY (SELECT * FROM tblname) TO xxx;.
- Table inheritance
- Not something we really need to deal with in the initial version, so long as it doesn't completely break (or we make sure it does for inheritance)
- What policy should be applied on the parent and child relations.
- idea: Also copy row-level policies from the parent tables.
- Foreign Key constraint
- Adding multiple modes for FK should be a separate, follow-on patch, it doesn't need to be in the initial version.
- idea: We can provide two modes: The first is filter-mode, the second is abort-mode.
- filter-mode: A normal mode. Row-level policy is evaluated earlier than user given condition, and returns false, if violated.
- abort-mode: A special internal mode. Row-level policy is evaluated after all the condition, and raises an error, if violated. The condition shall be evaluated earlier than row-level policy, the query has to be trusted. Such as queries in FK constraint.
- With RLS
- Without RLS
- The page.16 of LAPP/SELinux slides shows a pgbench comparison between pgsql-8.4.1 vs SE-PostgreSQL with RLS.
- It has 2-4% of performance tradeoff depending on database size.
- Note that SE-PostgreSQL implemented RLS in a different way than what is being proposed here; our goal is to get RLS support into core PG
- Integration with external security manager (eg: SELinux, SMACK)
- This will not be included in the initial support of RLS (unless we happen to get support for external security managers implemented first in PG)