Re: pg_upgrade segfaults when given an invalid PGSERVICE value

On 13-03-26 12:40 AM, Tom Lane wrote:
> Bruce Momjian <bruce(at)momjian(dot)us> writes:
>> On Mon, Mar 25, 2013 at 07:07:42PM -0400, Tom Lane wrote:
>>> Well, plan B would be to invent a replacement function that does have
>>> the ability to return an error message, but that seems like a lot of
>>> work for a problem that's so marginal that it wasn't noticed till now.
>>> (It's not so much creating the function that worries me, it's fixing
>>> clients to use it.)
>>>
>>> Plan C would be to redefine bogus value of PGSERVICE as not an error,
>>> period.
>
>> Given all of these poor options, is defining a PQconndefaults() as
>> perhaps out of memory or a service file problem really not better?
>
> Uh ... no. In the first place, what evidence have you got that those
> are (and will continue to be) the only two possible causes? In the
> second place, this still requires changing every client of
> PQconndefaults(), even if it's only to the extent of fixing their
> error message texts. If we're going to do that, I'd rather ask them
> to change to a more future-proof solution.
>

So to summarise:

Plan A: The first patch I attached for pg_upgrade + documentation
changes, and changing the other places that call PQconndefaults() to
accept failures on either out of memory, or an invalid PGSERVICE

Plan B: Create a new function PQconndefaults2(char * errorBuffer) or
something similar that returned error information to the caller.

Plan C: PQconndefaults() just ignores an invalid service but connection
attempts fail because other callers of conninfo_add_defaults still pay
attention to connection failures. This is the second patch I sent.

Plan D: Service lookup failures are always ignored by
conninfo_add_defaults. If you attempt to connect with a bad PGSERVICE
set it will behave as if no PGSERVICE value was set. I don't think
anyone explicitly proposed this yet.

Plan 'D' is the only option that I'm opposed to, it will effect a lot
more applications then ones that call PQconndefaults() and I feel it
will confuse users.

I'm not convinced plan B is worth the effort of having to maintain two
versions of PQconndefaults() for a while to fix a corner case.

> regards, tom lane
>
>