Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <75a2cb16-9e39-4123-de8e-b846c9b482f0@norrbonn.se>
Date:   Mon, 16 Jan 2023 09:59:27 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.4.2
Content-Language: en-US
To:     "Michael S. Tsirkin" <mst@redhat.com>,
        Jason Wang <jasowang@redhat.com>
Cc:     LKML <linux-kernel@vger.kernel.org>
From:   Jonas Bonn <jonas@norrbonn.se>
Subject: vhost-net
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hi,

I have a question about using vhost-net with an IFF_TUN device.  I'm 
uncertain about the behaviour I'm seeing which is:

i)  on RX, the descriptor contains no Ethernet header, which is what I 
was expecting
ii)  on TX, the first 14 bytes of the transmitted _IP_ packet are lost; 
if I prepend an extra 14 bytes (zeroes) before the IP packet and extend 
the packet length accordingly, then things appear to work as expected.

In vhost_net_build_xdp() it appears that the userspace packet data is 
copied verbatim to an XDP packet structure that assumes the presence of 
an ethernet header; as such, the IP header is copied into the ethernet 
header area.  I think this accounts for losing the first 14 bytes of the 
IP header...

If I set SO_SNDBUF to something less than INT_MAX, then the XDP path is 
bypassed and transmission of IP packets works.  This means that knowing 
the value of SO_SNDBUF becomes important in the userspace application in 
order to know whether an extra 14 bytes needs to be prepended to the IP 
packet... which is awkward, at best.

For an IFF_TUN device, should vhost-net not be adding an implicit 
ethernet header in _build_xdp()?  Can this be done without backward 
compatibility implications?

Thanks,
Jonas