Hi. Sorry for the slow follow-up. Stefan Hajnoczi <stefa...@gmail.com> writes:
> Yes, it's odd that QEMU changes make the issue go away but tcpdump > suggests the packet is not being sent from the bridge to the tap > device. Indeed. I really don't understand how to the two could possible interact! > e1000 and rtl8139 both use the same QEMU network subsystem code. I > don't see an obvious difference between the two. I wondered whether it could be some sort of checksum offloading which doesn't apply for virtio or rtl8139, but otherwise I agree it's extremely strange. > Since this issue only happens once in many QEMU runs are you sure that > -usbdevice tablet really makes the issue go away? I ran in a tight loop without it for around 500 iterations, and have done so again to confirm. Typically it fails within 20-40 iterations with -usbdevice tablet present. > Are you using ebtables? I know you mentioned disabling iptables and > it would be good to try the same for ebtables if you use it. We're normally using ebtables, but I completely flushed all the tables and set policy of ACCEPT for these tests to eliminate the possibility of bugs in my table create code. > In order to debug the host networking issue you may be able to use > ebtables/iptables LOG targets to collect information on how far > exactly the packets are getting. For example, you could try logging > all packets destined for the guest MAC address - and if the log > information includes the network interface you should see the packet > move between its source, the bridge, and the destination interface. I > have never tried this but it might work. I will have a play with this. Cheers, Chris.