Re: Shmem queue is not flushed if receiver is not yet attached

On Thu, Mar 17, 2022 at 3:13 AM Pavan Deolasee <pavan(dot)deolasee(at)gmail(dot)com> wrote:
> While testing on the current PG master, I noticed a problem between backends communicating over a shared memory queue. I think `shm_mq_sendv()` fails to flush the queue, even if `force_flush` is set to true, if the receiver is not yet attached to the queue. This simple fix solves the problem for me.
>
> On another note, `shm_mq.h` declares `shm_mq_flush()`, but I don't see it being implemented. Maybe just a leftover from the previous work? Though it seems useful to implement that API.

I think that this patch is basically correct, except that it's not
correct to set mqh_counterparty_attached when receiver is still NULL.
Here's a v2 where I've attempted to correct that while preserving the
essence of your proposed fix.

I'm not sure that we need a shm_mq_flush(), but we definitely don't
have one currently, so I've also adjusted your patch to remove the
dead prototype.

Please let me know your thoughts on the attached.

Thanks,

--
Robert Haas
EDB: http://www.enterprisedb.com