Re: reserved_superuser_connections tweak

Neil Conway <neilc@samurai.com> writes:
> --=-=-=
> There's no point counting the # of empty backend slots (which requires
> grabbing an LWLock) unless
>         (a) the # of reserved slots is > 0
>         (b) the current user is not a superuser
> Thus, we can sometimes get away without grabbing the lock

Good point.  OTOH, superuser() is not necessarily a cheap function
either.  If the user's pg_shadow row hasn't yet been loaded into the
SHADOWSYSID cache, it will provoke a catalog row fetch cycle, which will
cost *way* more than one measly LWLock.  I'm not sure whether that's
likely to be the case though --- have you checked where that cache entry
first gets loaded in a typical startup?

A more serious problem with this code, now that I look at it, is that it
is risking a catalog fetch cycle outside of any transaction.  That is a
Big No-No.

            regards, tom lane