[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20180418113058.GA9675@nautica>
Date: Wed, 18 Apr 2018 13:30:58 +0200
From: Dominique Martinet <asmadeus@...ewreck.org>
To: Jozsef Kadlecsik <kadlec@...ckhole.kfki.hu>
Cc: Florian Westphal <fw@...len.de>, Michal Kubecek <mkubecek@...e.cz>,
netdev@...r.kernel.org,
Marcelo Ricardo Leitner <marcelo.leitner@...il.com>,
Eric Dumazet <eric.dumazet@...il.com>
Subject: Re: tcp hang when socket fills up ?
Jozsef Kadlecsik wrote on Wed, Apr 18, 2018:
> Thanks for the testing! One more line is required, however: we have to get
> the assured bit set for the connection, see the new patch below.
I think it actually was better before. If I understand things correctly
at this point (when we get in the case TCP_CONNTRACK_SYN_RECV) we will
have seen SYN(out) SYN(in) SYNACK(out), but not the final ACK(in) yet.
Leaving old state as it was will not set the assured bit, but that will
be set on the next packet because old_state == new_state == established
at that point and the connection will really be setup then.
I don't think anything will blow up if we do either way, but strictly
speaking I'm more comfortable with the former.
I'll test the new patch regardless, I left work so can't reproduce
anymore but will yell tomorrow if it does explode ;)
> The tcp_conntracks state table could be fixed with introducing a new
> state, but that part is exposed to userspace (ctnetlink) and ugly
> compatibility code would be required for backward compatibility.
I agree a new state is more work than it is worth, I'm happy to leave it
as is.
--
Dominique Martinet | Asmadeus
Powered by blists - more mailing lists