On Fri, Jun 13, 2025 at 08:31:22PM +0530, Rahila Syed wrote:
> I am considering whether it would be better to avoid creating another DSM
> segment to track the DSA handle. Would it make more sense to track the
> DSAs in a separate dshash registry similar to DSM segments?
I don't know if it's better to manage 3 dshash tables than to manage 1 with
special entries for DSAs/dshash tables. There might be some small
trade-offs with each approach, but I haven't thought of anything too
worrisome...
> + /* Attach to existing DSA. */
> + dsa = dsa_attach(state->dsah);
> + dsa_pin_mapping(dsa);
> +
> + *found = true;
> + }
>
> Should this also consider the case where dsa is already mapped, to avoid
> the error on attaching to the DSA twice? IIUC, that would require calling
> dsa equivalent of dsm_find_mapping().
I wanted to find a way to do this, but AFAICT you can't. DSAs and dshash
tables are returned in backend-local memory, so if you lose that pointer, I
don't think there's a totally safe way to recover it. For now, I've
documented that GetNamedDSA()/GetNamedDSMHash() should only be called for a
given DSA/dshash once in each backend.
One other thing we could do is add a dsa_is_attached() function and then
ERROR if you try to reattach an already-attached DSA/dshash. I've done
this in the attached patch.
--
nathan
