On Sun, Aug 13, 2017 at 11:20 PM, Michael Paquier
<michael.paquier@gmail.com> wrote:
> Recent commit 8d98819 has added a missing permissoin check to lo_put()
> to make sure that the write permissions of the object are properly set
> before writing to a large object. When studying the problem, we bumped
> into the fact that it is possible to actually simplify those ACL
> checks and replace them by checks when opening the large object. This
> makes all the checks now in be-fsstubs.c happen in inv_api.c, which is
> where all the large object handling happens at a lower level. This
> way, it is also possible to remove the extra logic in place to have
> the permission checks happen only once.
>
> At the same time, we have bumped into two things:
> - ALLOW_DANGEROUS_LO_FUNCTIONS has been introduced in 1999, so it
> could be time to let it go. I have known of no place where this is
> actively used.
> - lo_import and lo_export on the backend have superuser checks. We
> could remove them and replace them with proper REVOKE permissions to
> allow administrators to give access to those functions.
>
> Attached is a set of 3 patches:
> - 0001 removes ALLOW_DANGEROUS_LO_FUNCTIONS
> - 0002 replaces the superuser checks with GRANT permissions
+1 for 0001 and 0002 in general, but I can't help noticing that they
lead to a noticeable worsening of the error messages in the regression
tests.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
