Re: xdp and fragments with virtio

[David Ahern <dsahern@gmail.com>]

On 5/16/18 1:24 AM, Jason Wang wrote:
> 
> 
> On 2018年05月16日 11:51, David Ahern wrote:
>> Hi Jason:
>>
>> I am trying to test MTU changes to the BPF fib_lookup helper and seeing
>> something odd. Hoping you can help.
>>
>> I have a VM with multiple virtio based NICs and tap backends. I install
>> the xdp program on eth1 and eth2 to do forwarding. In the host I send a
>> large packet to eth1:
>>
>> $ ping -s 1500 9.9.9.9
>>
>>
>> The tap device in the host sees 2 packets:
>>
>> $ sudo tcpdump -nv -i vm02-eth1
>> 20:44:33.943160 IP (tos 0x0, ttl 64, id 58746, offset 0, flags [+],
>> proto ICMP (1), length 1500)
>>      10.100.1.254 > 9.9.9.9: ICMP echo request, id 17917, seq 1,
>> length 1480
>> 20:44:33.943172 IP (tos 0x0, ttl 64, id 58746, offset 1480, flags
>> [none], proto ICMP (1), length 48)
>>      10.100.1.254 > 9.9.9.9: ip-proto-1
>>
>>
>> In the VM, the XDP program only sees the first packet, not the fragment.
>> I added a printk to the program (see diff below):
>>
>> $ cat trace_pipe
>>            <idle>-0     [003] ..s2   254.436467: 0: packet length 1514
>>
>>
>> Anything come to mind in the virtio xdp implementation that affects
>> fragment packets? I see this with both IPv4 and v6.
> 
> Not yet. But we do turn of tap gso when virtio has XDP set, but it
> shouldn't matter this case.
> 
> Will try to see what's wrong.
> 

I added this to the command line for the NICs and it works:

"mrg_rxbuf=off,guest_tso4=off,guest_tso6=off,guest_ecn=off,guest_ufo=off"

XDP program sees the full size packet and the fragment.

Fun fact: only adding mrg_rxbuf=off so that mergeable_rx_bufs is false
but big_packets is true generates a panic when it receives large packets.