On Thu, May 16, 2024 at 10:43 AM Thomas Munro <thomas.mu...@gmail.com> wrote: > Any chance you could test this version please Alexander?
Sorry, cancel that. v3 is not good. I assume it fixes the GSSAPI thing and is superficially better, but it doesn't handle code that calls twice in a row and ignores the first result (I know that PostgreSQL does that occasionally in a few places), and it's also broken if someone gets recv() = 0 (EOF), and then decides to wait anyway. The only ways I can think of to get full reliable poll()-like semantics is to do that peek every time, OR the complicated patch (per-socket-workspace + intercepting recv etc). So I'm back to v2.