On Fri, Aug 15, 2025 at 11:25:55AM -0500, Sami Imseih wrote:
> I noticed a few things where this behavior becomes very suspicious.
>
> For example, in pgss_startup_hook, every time startup_hook is run
> we take an exclusive LW lock. so, all backends now may be competing
> for that lock by nature of connection establishment.
I suspect there's enough overhead in connection establishment for
contention to be unlikely. In any case, I'm not aware of any complaints
about this.
> test_slru calls LWLockNewTrancheId inside that hook.
Hm. At first glance, that does seem bogus for EXEC_BACKEND builds. I
think the only side effect is extra tranche ID allocations and missing
tranche names, as SimpleLruInit() forgoes any shared memory initialization
in non-postmaster backends.
> So, this tells me that the caller needs to be aware of such caveats.
>
> I am thinking something like this:
>
> "Because this hook is executed by the postmaster and invoked by backends via
> EXEC_BACKEND, it is essential to ensure that any code intended to run only
> during postmaster startup is properly protected against repeated execution.
> This can be enforced by verifying !IsUnderPostmaster before invocation."
IMHO we should avoid talking about EXEC_BACKEND, etc. and instead make it
clear that hooks should be prepared to deal with concurrent invocations
from other backends. But taking a step back, I'm still not entirely clear
what this adds to the existing documentation, which is pretty direct about
the need for locking and how to avoid re-initializing.
--
nathan
