On Mon, Apr 7, 2025 at 10:05 AM Andres Freund <andres@anarazel.de> wrote:
> On 2025-04-07 09:41:25 -0700, Jacob Champion wrote:
> > Ah, you mean if the RPATH'd build doesn't have a flow, but the
> > globally installed one (with a different ABI) does? Yeah, that would
> > be a problem.
>
> That and more: Even if the RPATH libpq does have oauth support, the
> libpq-oauth won't necessarily be at the front of the global library search
> path. So afaict you'll often get a different libpq-oauth.
ldopen() should respect RPATH, though? Either way, I agree with
pushing the minor version into the name (or else deciding that we will
keep the ABI completely stable across minor version bumps; not sure I
want to guarantee that just yet).
> > We can do it however we want, honestly, especially since the ABI isn't
> > public/stable. I chose this way just to ease review.
>
> I found it rather confusing that libpq looks up a symbol and then sets
> libpq-oauth's symbol to a pointer in libpq's namespace.
Yeah, I think a one-time init call would make this nicer.
> > A future simplification could be to pull the use of the threadlock
> > back into libpq, and have it perform a one-time
> > dlopen-plus-Curl-initialization under the lock... That would also get
> > rid of the dlerror() thread safety problems. But that's an awful lot
> > of moving parts under a mutex, which makes me a little nervous.
>
> I still think we should simply reject at configure time if curl init isn't
> threadsafe ;)
Practically speaking, I don't think that's a choice we can make. For
example, RHEL won't have threadsafe Curl until 10.
> > Could you explain how this is related to .app bundles? I thought I was
> > just building a standard dylib.
>
> The other kind of bundles (what on earth apple was thinking with the naming
> here I don't know). Stuff liked with ld -bundle.
Ah, some new corner-case magic to learn...
> These days we use compiler flags
> that restrict function visibility of everything not annotated with
> PGDLLEXPORT.
Hm, I missed/forgot that. That is nice. Personally I like having a
single file document the exports, so I'll keep it that way for now
unless there are objections, but it's good to know it's not necessary.
Thanks,
--Jacob
