SE-PostgreSQL Updated Revision (r1460)

The patch set of SE-PostgreSQL and related stuff were updated (r1460).

[1/5] http://sepgsql.googlecode.com/files/sepostgresql-sepgsql-8.4devel-3-r1460.patch
[2/5] http://sepgsql.googlecode.com/files/sepostgresql-utils-8.4devel-3-r1460.patch
[3/5] http://sepgsql.googlecode.com/files/sepostgresql-policy-8.4devel-3-r1460.patch
[4/5] http://sepgsql.googlecode.com/files/sepostgresql-docs-8.4devel-3-r1460.patch
[5/5] http://sepgsql.googlecode.com/files/sepostgresql-tests-8.4devel-3-r1460.patch

I reviewed the patch set by myself, and updated the following items.
However, I would like other hackers to review the code in honesty.

SE-PostgreSQL need any volunteers to review and comment the patch set.
Please give us your support!

List of updates:
- Rebased to the latest CVS HEAD, which includes the column-level
privileges based on the SQL-standard.
(The previous r1425 conflicts in some points.)

- Security policy (sepostgresql-devel.pp) was updated to fit both of
Fedora 10 and rawhide. Test cases are also modified to care the new
security policy.

- Cleanup: NUM_SELINUX_CATALOG was replaced by lengthof() macro to
avoid code duplications.

- Cleanup: sepgsqlCheckEmbeddedProcedure() is renamed to
sepgsqlCheckProcedureInstall() due to its confusable naming.

- Add a new permission: db_procedure:{install}
It enables to prevent malicious user-defined functions are installed
as a part of operators, conversions, types and so on.
The default policy allows to install functions labeled as "sepgsql_proc_t"
only, as an implementation of these facilities.
Meanwhile, functions defined by unprivileged users are labeled as
"user_sepgsql_proc_t" in default, and it is not allowed to install as
an operator and so on.
If DBA want to install user-defined functions for the purpose, he has to
confirm its harmless and relabel it to "sepgsql_proc_t" at first.
In the previous revision, it checked "db_procedure:{execute}" here,
but it is not enough actually, because unprivilged user is allowed to
execute self defined function.

- Code revising: The previous revision always denied required permissions,
when the kernel does not define them within its security policy.
But it can make unexpected behavior when we work SE-PostgreSQL on
a system with legacy security policy which lacks a part of newly
added permissions.
The revised one simply allows actions when these are undefined.

- Fixbug: It required superfluous permissions when we try to update
"security_label" system column but it does not change anything actually.
For example:
UPDATE t SET security_label = security_label;
This query does not change security_label, so we don't need to check
"db_tuple:{relabelfrom}" permission here.
It is obvious we cannot know what tuples are actually relabeled on
sepgsqlExecScan(), so any permission checks for write-operations are
moved to sepgsqlHeapTuple(Insert|Update|Delete) hooks.

- Fixbug: when we update pg_largeobject system catalog by hand, it has
a possibility to create/drop specific largeobject, so we add a check
on "db_blob:{create drop}" when pg_largeobject.loid is modified by
UPDATE statement.
For example:
UPDATE pg_largeobject SET loid = loid::int + 10 WHERE loid = 1234;
It is theoretically same as dropping a largeobject with loid:1234 and
creating a largeobject with loid:1244.

- Fixbug: Tome Lane pointed out a matter when a whole-row-reference on
the relation with RTE_JOIN makes crash at the "Column-Level Privileges"
thread. This revision added a special care for the situation.
It recursively walks on refered JoinExpr and picks up its sources to
check permission to them.

- Code revising: T_SEvalItemRelation and T_SEvalItemAttribute nodes are
integrated into T_SelinuxEvalItem node. In the previous revision,
it simply chains all appeared tables and columns as a list of obsoleted
node on Query->pgaceItem. But it has a trend the length of list grows long.
T_SelinuxEvalItem contains required permissions on a table and an array of
permissions for columns. It enables to keep the length of the list minimum.
Related stuffs in sepgsql/proxy.c is also revised.
- addEvalRelation() / addEvalAttribute() enhanced to handle T_SelinuxEvalItem.
- Functions to handle inheritance tables and whole-row-reference are clearly
sorted out. expandEvalItemInheritance() handles inheritance tables, and
expandEvalItemWholeRowRefs() handles whole-row-reference.

- Add a hook: pgaceExecuteTruncate()
The previous revision checks permissions on truncated tables and tuples
on pgaceProcessUtility(), but this approach need to extract all the target
including cascaded ones, so it made code duplication.
The new hook is deployed on ExecuteTruncate() and delivers a list of already
opened relations with AccessExclusiveLock.
A new sepgsqlExecuteTruncate() checks needed permission on the hook.

- Cleanup: sepgsqlTupleName() always copied an identifier of tuple for audit
record into its internal buffer, and returns its pointer to caller.
But it is not necessary for most cases. It is revised to return a pointer
within given tuple to avoid useless strcpy(), if possible. So, its valid
duration is limited to the duration of tuple, but there is no real matter.
In addition, callers of sepgsqlTupleName() are cleaned up, because its
length of line tend to grow a bit long.

- Add a check: CREATE/ALTER FOREIGN DATA WRAPPER has a capability to load
a discretional shared library module, so it is necessary to check
db_database:{install_module} permission.

- Add source code comments: src/backend/security/sepgsql/avc.c

Thanks,
--
OSS Platform Development Division, NEC
KaiGai Kohei <kaigai(at)ak(dot)jp(dot)nec(dot)com>