Hi,
While working on an unrelated RLS patch, I noticed that equalPolicy() in
relcache.c doesn't compare the permissive field of RowSecurityPolicy. It
compares polcmd, hassublinks, policy_name, roles, qual, and
with_check_qual, but not permissive, so two policies that are identical
in every other respect except their PERMISSIVE/RESTRICTIVE designation
are reported as equal.
equalPolicy() is used (via equalRSDesc()) by RelationRebuildRelation()
to decide whether an open relation's existing row-security descriptor
can be kept as-is across a relcache rebuild, or must be treated as
changed. ALTER POLICY has no way to flip a policy's
PERMISSIVE/RESTRICTIVE designation, but DROP POLICY followed by CREATE
POLICY of the same name, roles, command, and quals, differing only in AS
PERMISSIVE/RESTRICTIVE, hits this exactly: the stale descriptor would be
kept, leaving the relcache out of sync with how the policy should now
combine with others (PERMISSIVE policies are ORed together; RESTRICTIVE
policies are ANDed with the rest).
I wasn't able to construct a simple SQL reproduction:
RelationRebuildRelation() (and thus this comparison) is only reached
while the relation is still open (refcount > 0) at the moment its
invalidation is processed. Closing and reopening the relation between
statements -- the usual behavior -- sidesteps the bug, and a second
session can't hold the relation open across the DROP/CREATE either,
since both need AccessExclusiveLock. This looks analogous to
ab6d1cd26eb, which fixed the same kind of omission in this function for
the USING qual and likewise shipped without a test.
Patch attached.
Regards,
Andreas Lind
