Re: Problem with PostgreSQL 18 and semaphores

[Martin Husemann <martin%duskware.de@localhost>]

On Sun, Oct 26, 2025 at 11:59:17AM +0100, Marc Baudoin wrote:
> So I believe those should be raised (for 11.0 and 10.2) because
> PostgreSQL should be able to run out of the box on NetBSD.

It is not that simple and Postgres is an exception and not run on most
NetBSD systems.

The problem is that the allocation is static and requires wired kernel
memory. Just bumping up the numbers will waste lots of this resourse
(which is quite scare on some architectures).

I think it is fine for the admin to bump values higher (via the sysctls
Manuel mentioned, and save them in /etc/sysctl.conf) once they run into
the problem.

Bumping the limits up automatically (e.g. based on available KVA +
installed RAM) would be cool, but we should move to a more dynamic
allocation scheme then - which is not rocket science, but unlikely to
happen for 11.0 or 10.2.

What we could do short-term is to modify the postgres message #ifdef
__NetBSD__ and point to a man page describing the details. The sysctls
seem to not be documented besides their own description (sysctl -d) and
sysctl(7). Maybe sem_open(3) should get a paragraph about the issue and
suggestions how to work around when the system hits the limit?

Martin