2013/9/4 Tom Lane <tgl@sss.pgh.pa.us>:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Wed, Sep 4, 2013 at 10:50 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>> Robert Haas <robertmhaas@gmail.com> writes:
>>>> Right. IMHO, this new feature should be similarly simple: when an
>>>> unprivileged user references a table, treat that as a reference to a
>>>> leakproof view over the table, with the RLS qual injected into the
>>>> view.
>
>>> And for insert/update/delete, we do what exactly?
>
>> The same mechanism will prevent UPDATE and DELETE from seeing any rows
>> the user shouldn't be able to touch.
>
> No, it won't, because we don't support direct update/delete on views
> (and if you look, you'll notice the auto-updatable-view stuff doesn't
> think a security-barrier view is auto-updatable).
>
> AFAICT, to deal with update/delete the RLS patch needs to constrain order
> of qual application without the crutch of having a separate level of
> subquery; and it's that behavior that I have zero confidence in, either
> as to whether it works as submitted or as to our odds of not breaking it
> in the future.
>
Are you suggesting to rewrite update / delete statement to filter out
unprivileged rows from manipulation?
Yes. I also thought it is a simple solution that does not need additional
enhancement to allow update / delete to take sub-query on top of reader
side plan.
For example, if security policy is (t1.owner = current_user) and the given
query was "UPDATE t1 SET value = value || '_updated' WHERE value like '%abc%'",
this query may be able to rewritten as follows: UPDATE t1 SET value = value || '_updated' WHERE tid = ( SELECT tid
FROMt1 WHERE t1.owner = current_user ) AND value like '%abc%';
This approach makes implementation simple, but it has to scan the
relation twice, thus its performance it not ideal, according to the
past discussion.
Thanks,
--
KaiGai Kohei <kaigai@kaigai.gr.jp>