Re: Remove header lock BufferGetLSNAtomic() on architectures with 64 bit atomic operations

Hi,

On 2025-11-24 00:10:03 +0100, Andreas Karlsson wrote:
> Andres pointed out this possible optimization on Discord so I hacked up a
> quick patch which avoids taking a lock when reading the LSN from a page on
> architectures where we can be sure to not get a torn value. It is always
> nice to remove a lock from a reasonably hot code path.

Nice.

> static inline XLogRecPtr
> PageXLogRecPtrGet(PageXLogRecPtr val)
> {
> - return (uint64) val.xlogid << 32 | val.xrecoff;
> + return val;
> }
>
> #define PageXLogRecPtrSet(ptr, lsn) \
> - ((ptr).xlogid = (uint32) ((lsn) >> 32), (ptr).xrecoff = (uint32) (lsn))
> + ((ptr) = (lsn))
> +
> +#else
> +
> +static inline XLogRecPtr
> +PageXLogRecPtrGet(volatile PageXLogRecPtr val)
> +{
> + return (val << 32) | (val >> 32);
> +}

A volatile on a non-pointer won't do you much good, I'm afraid. You need to
make sure that the underlying value is read as a single 8 byte read, I don't
see how this guarantees that, unfortunately.

Greetings,

Andres Freund