On Fri, Aug 15, 2025 at 09:17:22AM -0500, Sami Imseih wrote:
> While working on a related area, I noticed that shmem_startup_hook
> is called twice under EXEC_BACKEND.
>
> The first call occurs in CreateSharedMemoryAndSemaphores() during
> postmaster startup (!IsUnderPostmaster), which is expected.
>
> The second call happens in AttachSharedMemoryStructs() (when
> EXEC_BACKEND is defined), and this occurs in normal backends
> (IsUnderPostmaster).
>
> The second call does not seem correct. The startup hook that should
> only run during
> postmaster initialization, AFAIK.
>
> Blame shows that this change was introduced in commit 69d903367c,
> but I could not determine the rationale from the discussion,
> so it may have been an oversight.
I think the startup hook must run in each backend for EXEC_BACKEND, else we
won't properly initialize pointers to shared memory in that case, right?
I added the following wording in commit 964152c:
+ <literal>shmem_startup_hook</literal> provides a convenient place for the
+ initialization code, but it is not strictly required that all such code
+ be placed in this hook. Each backend will execute the registered
+ <literal>shmem_startup_hook</literal> shortly after it attaches to shared
+ memory. Note that add-ins should still acquire
+ <function>AddinShmemInitLock</function> within this hook, as shown in the
+ example above.
IIUC commit 69d9033 actually makes that inaccurate for the non-EXEC_BACKEND
case. Presumably this is okay because we don't need to re-initialize
pointers to shmem when forking. I must've missed this change when updating
the documentation.
--
nathan
