Hi Konstantin,
On Thu, Aug 30, 2018 at 11:27:31AM -0700, Michael Paquier wrote:
> It seems to me that you are right here, "path" points to
> pg_replslot/$SLOTNAME/state which is a file so the fsync is incorrect.
> I am not sure that we'd want to publish fsync_parent_path out of fd.c
> though, so perhaps we could just save the slot path in a different
> variable and use it?
I have spent more time on this bug, and the code path you have pointed
at is the only one having such an issue. Attached is a patch to fix the
problem. It includes the sanity checks I have used to check all code
paths calling fsync_fname() for both the frontend and the backend code.
The checks will not be included in the final fix, still they look useful
so I am planning to start a new thread on the matter as perhaps other
folks have more and/or better ideas.
--
Michael
