Re: protect dll lib initialisation against any exception, for 8.5

Pavel Stehule <pavel(dot)stehule(at)gmail(dot)com> writes:
> attached patch allows raising exception from _PG_init function as was
> discussed before.

I fooled around with this and came up with the attached improved
version, which allows reporting the full error status. However,
after thinking some more I feel that this is probably a cure worse
than the disease. If we simply leave the code as it stands, an
elog(ERROR) in an init function doesn't corrupt dfmgr.c's internal list,
which is what I had been fearing when I complained about the issue.
The worst that happens is that we leave the library loaded and leak
a little bit of memory. Unloading the library, as the patch does,
could easily make things worse not better. Consider the not-unlikely
case that the library installs itself in a few callback hooks and
then fails. If we unload the library, those hooks represent
*guaranteed* core dumps on next use. If we don't unload, the hook
functions might or might not work too well --- presumably not everything
they need has been initialized --- but it's hard to imagine an outcome
that's worse than a guaranteed core dump.

So I'm thinking this is really unnecessary and we should leave well
enough alone.

regards, tom lane