Re: ReplicationSlotRelease() crashes when the instance is in the single user mode

On Monday, February 17, 2025, Hayato Kuroda (Fujitsu) <
kuroda(dot)hayato(at)fujitsu(dot)com> wrote:

>
> backend> SELECT pg_create_physical_replication_slot(slot_name :=
> 'physical_slot', immediately_reserve := true);

Since this function releases the slot when it returns, re-acquisition, even
by the same backend, must always re-associate MyProcPid to the named slot.

>
> [1]:
> ```
> --- a/src/backend/replication/slot.c
> +++ b/src/backend/replication/slot.c
> @@ -599,7 +599,7 @@ retry:
> SpinLockRelease(&s->mutex);
> }
> else
> - active_pid = MyProcPid;
> + s->active_pid = active_pid = MyProcPid;
> LWLockRelease(ReplicationSlotControlLock);
>
> /*
> ```
>

This, but you cannot modify the slot without holding the spinlock.

I’d probably add an assert that the existing state of s->active_pid is
either 0 or MyProcPid already. In single-user mode it mustn’t, really
cannot, be anything else. But the failure here is because the SQL function
does a slot release; there are probably other reasonable paths where the
assignment of MyProcPid during slot creation is retained and encountered
during a subsequent slot acquire.

David J.