Re: Replication to Postgres 10 on Windows is broken

On 2017-08-06 18:04:49 -0400, Tom Lane wrote:
> Andres Freund <andres(at)anarazel(dot)de> writes:
> > Here's a prototype patch implementing what Tom outlined.
>
> This bit is flat wrong:
>
> - int io_flag;
> + int io_flag = WL_POSTMASTER_DEATH | WL_LATCH_SET;
>
> io_flag has to be *just* the I/O condition, because we use it in a test
> after the WaitLatchOrSocket call.

Hm, right. Wouldn't be particularly consequential, but... I'd actually
consider just removing the if around
/* If socket is ready, advance the libpq state machine */
if (rc & io_flag)
status = PQconnectPoll(conn->streamConn);
the only thing that protects us against is calling PQconnectPoll() when
the latch has been set. Hardly problematic.

> > Anybody have an opinion about adding ifs for WL_SOCKET_CONNECTED to
> > !win32 implementations rather than redefining it to WL_SOCKET_WRITEABLE?
>
> I fear it would complicate matters greatly, because you'd have to figure
> out which of the two flags to signal back after detecting socket writable.
> I think defining it as equal to WL_SOCKET_WRITEABLE is fine.

Well, I'd have said, signal the one(s) back that have been
requested. But I'm ok with the current state, adding a bunch of
pointless branches didn't strike me as worthwhile...

Greetings,

Andres Freund