Hi Daniel > Hi all, > > Dmitry and I, we thought about a possible implementation allowing the > l2/l3 to coexists. > > The idea is assuming the l3 network namespaces are the leaf in the l2 > namespace hierarchy tree. By default, init process is l2 namespace. From > a layer 3, it is impossible to do a new network namespace unshare. > > All the configuration is done into the l2 namespace. When a l3 is > created a new IP address should be created into the l2 namespace and > "pushed" into the l3. When the l3 dies, the IP is pulled to its parent, > aka the l2. In order to ensure security into the l3, the NET_ADMIN > capability is lost when doing unsharing for l3. > There is no extra code for socket virtualization. It is a common part. > > How to setup a l3 namespace ? > ----------------------------- > > 1 - setup a new IP address in l2 namespace > 2 - create a l3 namespace > 3 - specific socket ioctl to "push" the IP address from the l2 > namespace to the newly created l3 namespace This means that there is some kind of identifier for the l3 namespace, right? > > The l2 lose visibility on the IP address and l3 gains visibility on the > IP address. A ifconfig or a ip command shows only the IP address > assigned to the namespace. Loopback address is always visible. Hmm.... I've been thinking about this, and I think this OK from the sockets point of view, i.e. binds() in l2 lose visibility to the new l3 address. There is a concern for a potential race here though. However, it would be really nice to be able to see l3 namespace addresses in the parent l2 tagged in some way. > > How to handle outgoing traffic ? > -------------------------------- > > The bind must be checked with the IP addresses belonging to the l3 > namespace and with all the derivative addresses (multicast, broadcast, > zero net, loopback, ...). > > The IP addresses will rely on aliased IP address. The source address > must be filled with the IP address belonging the l3 namespace when not > set. This is a trivial operation, because we know which IP addresses are > assigned to the l3 namespace. Can you provide a little more info? > > When the route are resolved, the l3 namespace switch the its parent, > that is to say the l2 namespace, and the virtualization follows its > normal path. > > How to handle incoming traffic ? > -------------------------------- > > Because we can have several sockets listening on the same > INADDR_ANY:port, we must find the network namespace associated with the > destination IP address. > For unicast, this is a trivial operation, because that can be checked > with the assigned IP address again. For broadcast and multicast, some > extra work should be done in order to store the namespaces which are > listening on a broadcast address. As soon as the namespace is found, we > switch to it. This can be done with netfilters. The problem is with multicasts. Multicast groups are joined on the interface bases. Every socket that bound *:multicast_port will receive multicast traffic once a single app joined the group. Since l3 namespaces don't have share the conceptual interface, theoretically, all l3 namespaces should receive multicast traffic. > > Routes and co. > -------------- > > - Routes: they are not isolated, each l3 namespace can see all the > routes from the other namespaces. That allows the routing engine to see > all the routes and choose the loopback when two network namespaces in > the same host try to communicate. > > - Cache: the routing cache must be isolated, otherwise the socket > isolation will not work. The l3 namespace code does not impact the l2 > namespace code and route cache isolation is a common part if the l3 > namespace switching is done in the right place. > > > Dmitry has posted the l2 namespace relying on the net namespace empty > framework, I will post the l3 namespace relying on the l2 namespace > today or tomorrow. > Looking forward to it. Thanks -vlad