Hi Dmitry, On Wed, Mar 15, 2017 at 12:45:54AM +0300, Dmitry Sivachenko wrote: > I committed your patch to FreeBSD ports.
I was just reported an undesired side effect of this patch with smtp in clear without proxy-proto :-( The problem is that we're using the CONNECTED flag to indicate whether we've just detected the transition from pending handshake to connected or not. And the data layer also needs to see the transition so we can't set it too early. Thus in order to fix Matthias' issue I'm tempted by extending the CONN_STATE flag to cover all handshakes, but I'm scared to break complex health checks by calling them multiple times when multiple layers are stacked (eg: ssl + send_proxy). The root cause of the problem is the fact that we don't cover the handshakes completion when deciding to wake the data layer up, and that these ones are supposed to (sometimes) complete a connection establishment (eg: Matthias' case where the successful connect() didn't leave room for another event to report this). I *think* we can use as a universal event the absence of any handshake, any pending L4/L6 pending connection and the absence of the CO_FL_CONNECTED flag, indicating we're just finishing a connection validation. I need to verify if it would work both for checks and regular connections, and on both sides, in a wide variety of combinations. I also don't feel much confident by the fact that the stream interface layer validates the connection establishment without checking for pending handshakes, so such a change requires extreme care :-/ I'm now working on this (I didn't expect to do this today) but I already feel like I'm pulling one thread and that the whole ball is coming with it. Let's hope that it will end up with all these design bugs being smashed :-/ In your case, if you get bad reports it might be better to go back to the previous package with the temporary revert. But I hope to have good news today so that you don't have to emit two package updates. Regards, Willy
