Re: logical replication launcher crash on buildfarm

On 2017-03-16 09:27:59 -0400, Robert Haas wrote:
> On Thu, Mar 16, 2017 at 5:13 AM, Petr Jelinek
> <petr(dot)jelinek(at)2ndquadrant(dot)com> wrote:
> > Hmm now that you mention it, I remember discussing something similar
> > with you last year in Dallas in regards to parallel query. IIRC Windows
> > should not have this problem but other systems with EXEC_BACKEND do.
> > Don't remember the details though.
>
> Generally, extension code can't use bgw_main safely, and must use
> bgw_library_name and bgw_function_name. But bgw_main is supposedly
> safe for core code.

I indeed think it's not safe, and it's going to get less and less safe
on windows (or EXEC_BACKEND). I don't think we can afford to disable
ASLR in the long run (I indeed supect that'll just be disallowed at some
point), and that's the only thing making it safe-ish in combination with
EXEC_BACKEND.

> If it's not even safe there, then I guess we should remove it entirely
> as a useless foot-gun.

I indeed think that's the right consequence. One question is what to
replace it with exactly - are we guaranteed we can dynamically lookup
symbols by name in the main binary on every platform? Alternatively we
can just hardcode a bunch of bgw_function_name values that are matched
to specific functions if bgw_library_name is NULL - I suspect that'd be
the easiest / least worrysome portability-wise.

Greetings,

Andres Freund