Re: locked reads for atomics

Hi,

On 2024-02-23 14:58:12 -0600, Nathan Bossart wrote:
> +/*
> + * pg_atomic_write_membarrier_u32 - write with barrier semantics.
> + *
> + * The write is guaranteed to succeed as a whole, i.e., it's not possible to
> + * observe a partial write for any reader. Note that this correctly interacts
> + * with both pg_atomic_compare_exchange_u32() and
> + * pg_atomic_read_membarrier_u32(). While this may be less performant than
> + * pg_atomic_write_u32() and pg_atomic_unlocked_write_u32(), it may be easier
> + * to reason about correctness with this function in less performance-sensitive
> + * code.
> + *
> + * Full barrier semantics.
> + */

The callout to pg_atomic_unlocked_write_u32() is wrong. The reason to use
pg_atomic_unlocked_write_u32() is for variables where we do not ever want to
fall back to spinlocks/semaphores, because the underlying variable isn't
actually shared. In those cases using the other variants is a bug. The only
use of pg_atomic_unlocked_write_u32() is temp-table buffers which share the
data structure with the shared buffers case.

Greetings,

Andres Freund