From mboxrd@z Thu Jan 1 00:00:00 1970 From: Cong Wang Subject: Re: [PATCH net-next] net: preserve sock reference when scrubbing the skb. Date: Thu, 28 Jun 2018 16:18:47 -0700 Message-ID: References: <20180625155610.30802-1-fbl@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Cc: Linux Kernel Network Developers , Eric Dumazet , Paolo Abeni , David Miller , Florian Westphal , NetFilter To: Flavio Leitner Return-path: Received: from mail-pl0-f65.google.com ([209.85.160.65]:36780 "EHLO mail-pl0-f65.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932883AbeF1XTB (ORCPT ); Thu, 28 Jun 2018 19:19:01 -0400 In-Reply-To: <20180625155610.30802-1-fbl@redhat.com> Sender: netdev-owner@vger.kernel.org List-ID: On Mon, Jun 25, 2018 at 8:59 AM Flavio Leitner wrote: > XPS breaks because the queue mapping stored in the socket is not > available, so another random queue might be selected when the stack > needs to transmit something like a TCP ACK, or TCP Retransmissions. > That causes packet re-ordering and/or performance issues. Now let me look at the XPS part, a key question first: By queue mapping stored in socket, you mean sk_tx_queue_get(), which is only called in __netdev_pick_tx(), and of course even before hitting qdisc layer. However, veth device orphans the skb inside its veth_xmit(), (dev_forward_skb()), which is after going through qdisc layer. So, how could the skb_orphan() called _after_ XPS break XPS? We are talking about a simple netns-to-netns case, so XPS won't be hit again once leaves it. Another _dumb_ question: veth is virtual device, it has literally no queues, I know technically there is a queue for installing qdisc. So, why does even queue mapping matters here???