Re: libpq's multi-threaded SSL callback handling is busted

Jan Urbański writes:

> I did some more digging on bug
> http://www.postgresql.org/message-id/CAHUL3dpWYFnUgdgo95OHYDQ4kugdnBKPTjq0mNbTuBhCMG4xvQ@mail.gmail.com
> which describes a deadlock when using libpq with SSL in a multi-threaded
> environment with other threads doing SSL independently.
>
> [reproducing instructions]
>
> I posit we should remove all CRYPTO_set_*_callback functions and associated
> cruft from libpq.
>
> I could submit a patch to get rid of the crazy CRYPTO_*_callback dance in
> libpq, but at the very least this will require a warning in the release notes

Attached is a patch doing just that.

> I would very much like to have this change back-patched, since setting and
> resetting the callback makes using libpq in a threaded OpenSSL-enabled app
> arguably less safe than if it didn't use any locking.

Also attached is a patch for 9.4 and all previous supported releases, which is
the same thing, but adjusted for when we didn't have a separate fe-secure.c and
fe-secure-openssl.c

If committed, this change will warrant a notice in the release notes. I could
try drafting it, if that'd be helpful.

Cheers,
Jan