Horton

Re: pgsql: Teach DSM registry to ERROR if attaching to an uninitialized ent

Nathan Bossart <nathandbossart@gmail.com>

View thread View on postgr.es
From: Nathan Bossart <nathandbossart@gmail.com>
To: Robert Haas <robertmhaas@gmail.com>
Cc: PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Date: 2025-11-24T21:25:51Z
Lists: pgsql-hackers 
On Mon, Nov 24, 2025 at 04:02:18PM -0500, Robert Haas wrote:
> I haven't done a full review of this and I'm not sure whether you want
> me to spend more time on it...

I'd appreciate an eyeball check if you have the time.

> I'm guessing that the reason why that doesn't easily work is
> because you're relying on those locks to prevent multiple backends
> from doing the same initialization?

For GetNamedDSMSegment(), I bet we could avoid needing a PG_CATCH by taking
the DSMRegistryLock exclusively when accessing the registry.  But for
GetNamedDSA() and GetNamedDSHash(), we want to keep the half-initialized
entry so that we don't leak LWLock tranche IDs.  I initially thought that
might be okay, but if every backend is retrying, you could quickly run out
of tranche IDs.

-- 
nathan

Commits

  1. Teach DSM registry to retry entry initialization if needed.

  2. Revert "Teach DSM registry to ERROR if attaching to an uninitialized entry."

  3. Teach DSM registry to ERROR if attaching to an uninitialized entry.