Altering view ownership doesn't work ...

... because nowhere does it update the "checkAsUser" fields in the
view's query to be the OID of the new owner. This means that
permission checks about whether the view can access its underlying
tables will still be done as the old owner. An example:

regression=# create user u1;
CREATE ROLE
regression=# create user u2;
CREATE ROLE
regression=# \c - u1
You are now connected to database "regression" as user "u1".
regression=> create table t1(f1 int);
CREATE TABLE
regression=> create view v1 as select * from t1;
CREATE VIEW
regression=> grant select on v1 to u2;
GRANT

-- at this point u2 can select from v1 but not directly from t1

regression=> \c - postgres
You are now connected to database "regression" as user "postgres".
regression=# alter table v1 owner to u2;
ALTER TABLE
regression=# \c - u2
You are now connected to database "regression" as user "u2".
regression=> select * from v1;
f1
----
(0 rows)

-- this is WRONG, u2 should not have any ability to select from t1

The same problem applies to all rules, really, not only a view's
ON SELECT rule.

This is particularly bad because pg_dump is relying heavily on
ALTER OWNER these days. After a dump/restore, it is likely that
every view's "original owner" will be a superuser, and thus that
all permission checking is effectively disabled for accesses
from views. It wouldn't be too much of a stretch to call that
a security loophole.

I can think of two basic ways to fix this:

1. Add a bunch of code to ALTER OWNER to update every rule attached to
the target table.

2. Run setRuleCheckAsUser during rule load rather than rule store.

#2 is a lot simpler, and would fix the problem for existing broken rules
whereas #1 would not, so I'm kind of inclined to go with that. I doubt
there'd be any meaningful performance hit --- parsing the stored form
of a rule is relatively expensive anyway, so we cache the results.

Comments?

regards, tom lane