Re: [Qemu-devel] virtio-net & dhcp & udp checksum

[Michael S. Tsirkin]

On Tue, Sep 06, 2016 at 09:00:30PM +1000, Alexey Kardashevskiy wrote:
> On 06/09/16 20:03, Michael S. Tsirkin wrote:
> > On Tue, Sep 06, 2016 at 02:32:27PM +1000, Alexey Kardashevskiy wrote:
> >> Hi!
> >>
> >> I am trying DHCP between 2 guests. So I am running first guest with:
> >>
> >> -netdev tap,id=TAP0,helper=/home/aik/qemu-bridge-helper \
> >> -device "virtio-net-pci,id=vnet0,mac=C0:41:49:4b:ee:ee,netdev=TAP0"
> >>
> >> and second one with:
> >>
> >> -netdev tap,id=TAP0,vhost=on,helper=/home/aik/qemu-bridge-helper \
> >> -device "virtio-net-pci,id=vnet0,mac=C0:41:49:4b:00:01,netdev=TAP0" \
> >>
> >>
> >> Both tap are connected to br0 on the host:
> >>
> >> address@hidden:~$ brctl show
> >> bridge name     bridge id               STP enabled     interfaces
> >> br0             8000.fe397c73cecc       no              tap0
> >>                                                         tap1
> >>
> >> Both guests are debian8 with v4.7 kernel, one is running Dnsmasq version
> >> 2.72, the other - isc-dhclient-4.3.1.
> >>
> >> The very first response from dnsmasq has a bad UDP checksum:
> >>
> >> 04:19:04.946754 c0:41:49:4b:ee:ee > c0:41:49:4b:00:01, ethertype IPv4 
> >> (0x0800)
> >> , length 346: (tos 0xc0, ttl 64, id 60635, offset 0, flags [none], proto 
> >> UDP (
> >> 17), length 332)
> >>     192.168.1.250.67 > 192.168.1.1.68: [bad udp cksum 0x8595 -> 0x6e44!] 
> >> BOOTP
> >> /DHCP, Reply, length 304, xid 0x38e6b51c, Flags [none] (0x0000)
> >>
> >> 0x8595 looks like a UDP header checksum. Unlike dhclient (which uses
> >> PF_PACKET), dnsmasq seems to use AF_INET and DGRAM so I am wondering what
> >> exactly should do this checksum calculations in this case and why it does
> >> not do this?
> > 
> > Receiver should - the packet is clearly marked as such.
> 
> Where is that mark exactly? UDP header (unlikely)? IP header (cannot see
> it)? DHCP?

It's returned in the aux data. See linux commit below.

> 
> > Of course old dhclient ignores this flag.  I think Debian used to
> > carry a patch to make it take it into account.
> 
> So seeing bad UDP checksum message in tcpdump of the _source_ (dnsmasq's
> tap) side is ok?

Yes - no one bothered teaching tcpdump about this commit:

    commit 8dc4194474159660d7f37c495e3fc3f10d0db8cc
    Author: Herbert Xu <address@hidden>
    Date:   Sun Feb 4 23:31:32 2007 -0800
        [PACKET]: Add optional checksum computation for recvmsg


> >>
> >> I read the old discussion at
> >> http://www.mail-archive.com/address@hidden/msg41958.html
> >>
> >> but it seems that in my case the broken thing is dnsmasq (which is hard to
> >> believe). Running the dnsmasq's guest with
> >>
> >> -device virtio-net-pci,id=vnet0,csum=off,...
> >>
> >> fixes the problem.
> >>
> >> What is broken now? Thanks.
> > 
> > Maybe debian dropped the patch from dhcp?
> > http://marc.info/?l=kvm&m=129347023113779
> > 
> > Or you can use the host-side workaround using a checksum rule.
> 
> Sure, I am just trying to educate myself and understand what to expect and
> what is broken :)

If you have the time, try adding the patch to dhcp in debian.
Maybe it was ubuntu that was carrying it? You can try that ...


> 
> 
> -- 
> Alexey