Stefan Hajnoczi, le Wed 06 Mar 2013 13:29:37 +0100, a écrit : > On Tue, Mar 05, 2013 at 05:35:10PM +0100, Samuel Thibault wrote: > > The reason why IPv6 does not work when using -net socket,mcast=foo is > > that since qemu explicitly sets IP_MULTICAST_LOOP to 1, it receives its > > own frames. When the IPv6 stack performs duplicate addresse detection > > (DAD) through a multicasted announce, it receives its own announcement, > > and thus believes another machine has the same address. > > The reason for IP_MULTICAST_LOOP is to allow QEMU processes running on > the same host to communicate with each other.
Sure, I've seen the comment, I wasn't suggesting to drop that :) > > AIUI, on a real physical network network boards do not receive the > > multicasts they send, so the issue does not happen. Perhaps some boards > > even filter out any frame with its own MAC as source, eliminating the > > issue altogether. > > > > As a result, we should probably perform this kind of dropping, I'm just > > wondering at which level that would be preferable. > > > > - We could do that in qemu_send_packet_async_with_flags, thus fixing the > > issue for all kinds of frame transporters. > > > > - Or we could do that for the only case that matters, mcast, in > > net_socket_send_dgram (which will thus do it for the unicast udp case > > too). > > > > What do people think about it? > > We should fix the layer that introduces the problem. Therefore I think > the fix needs to be net/socket.c. Ok. > Unfortunately net/socket.c does not have the concept of a link-layer > address, so we cannot easily filter out multicast packets coming from > our NIC's address. > > Are you aware of a way to filter out just the packets sent by *this* > process? I haven't seen any in the Linux source code. One thing that should work, however, is to use recvfrom, and drop whatever comes from our sockname. Samuel
