Re: Atomics in localbuf.c

Andres Freund <andres(at)anarazel(dot)de> wrote:

> On March 5, 2020 9:21:55 AM PST, Antonin Houska <ah(at)cybertec(dot)at> wrote:
> >What's the reason to use pg_atomic...read_...() and
> >pg_atomic...write_...()
> >functions in localbuf.c?
> >
> >It looks like there was an intention not to use them
> >
> >https://www.postgresql.org/message-id/CAPpHfdtfr3Aj7xJonXaKR8iY2p8uXOQ%2Be4BMpMDAM_5R4OcaDA%40mail.gmail.com
> >
> >but the following discussion does not explain the decision to use them.
>
> Read/write don't trigger locked/atomic operations. They just guarantee that
> you're not gonna read/write a torn value. Or a cached one. Since
> local/shared buffers share the buffer header definition, we still have to
> use proper functions to access the atomic variables.

Sure, the atomic operations are necessary for shared buffers, but I still
don't understand why they are needed for *local* buffers - local buffers their
headers (BufferDesc) in process local memory, so there should be no concerns
about concurrent access.

Another thing that makes me confused is this comment in InitLocalBuffers():

/*
* Intentionally do not initialize the buffer's atomic variable
* (besides zeroing the underlying memory above). That way we get
* errors on platforms without atomics, if somebody (re-)introduces
* atomic operations for local buffers.
*/

That sounds like there was an intention not to use the atomic functions in
localbuf.c, but eventually they ended up there. Do I still miss something?

--
Antonin Houska
Web: https://www.cybertec-postgresql.com