Replace known_assigned_xids_lck by memory barrier - Mailing list pgsql-hackers

Hello everyone and Tom.

Tom, this is about your idea (1) from 2010 to replace spinlock with a
memory barrier in a known assignment xids machinery.

It was mentioned by you again in (2) and in (3) we have decided to
extract this change into a separate commitfest entry.

So, creating it here with a rebased version of (4).

In a nutshell: KnownAssignedXids as well as the head/tail pointers are
modified only by the startup process, so spinlock is used to ensure
that updates of the array and head/tail pointers are seen in a correct
order. It is enough to pass the barrier after writing to the array
(but before updating the pointers) to achieve the same result.

Best regards.

[1]:
https://github.com/postgres/postgres/commit/2871b4618af1acc85665eec0912c48f8341504c4#diff-8879f0173be303070ab7931db7c757c96796d84402640b9e386a4150ed97b179R2408-R2412

[2]: https://www.postgresql.org/message-id/flat/1249332.1668553589%40sss.pgh.pa.us#19d00eb435340f5c5455e3bf259eccc8

[3]: https://www.postgresql.org/message-id/flat/1225350.1669757944%40sss.pgh.pa.us#23ca1956e694910fd7795a514a3bc79f

[4]:
https://www.postgresql.org/message-id/flat/CANtu0oiPoSdQsjRd6Red5WMHi1E83d2%2B-bM9J6dtWR3c5Tap9g%40mail.gmail.com#cc4827dee902978f93278732435e8521