Re: [libvirt] [PATCHv2 0/9] Let libvirt manage a bridge's FDB

On Tue, Dec 02, 2014 at 12:08:30PM -0500, Laine Stump wrote: > (Everyone - see the request for opinions/ideas towards the bottom) > > The idea behind these patches is the following: > > 1) most virtual machines only have a single MAC address behind each > interface, and that MAC address is known by libvirt. > > 2) If we (i.e. libvirt) manually add an entry to the bridge's > forwarding database (fdb) for the MAC address associated with a port > on the bridge, we can turn off learning and unicast_flooding for that > port. > > 3) kernels starting with 3.15 (and actually working correctly starting > in kernel 3.17) will notice that all of a bridge's ports have flood > and learning turned off, and in that case will turn off promiscuous > mode on all ports. If all but one of the ports have flood/learning > turned off, then promiscuous will be turned off on that port (and left > on for all the other ports) > > 4) When (4) can be done, there is a measurable performance > advantage. It can also *kind of* help security, as it will prevent a > guest from doing anything useful if it changes its MAC address (but > won't prevent the guest from *sending* packets with a spoofed MAC > address). > > NB: These only work with a fixed MAC address, and no vlan tags set in > the guest. Support for both of those will be coming. So just to be clear on the functional differences between the impls With the current impl 1. The bridge has a table mapping MAC <-> TAP devices which is initially empty 2. Whenever a TAP device sends a packet, it causes an entry to be added to the MAC <-> TAP table 3. When the bridge needs to forward a packet with a MAC - If MAC is unknown, it is sent to all TAP devices - Else it is send to the specific TAP device 4. The physical device is always promiscuous to see all traffic on the LAN An result of point 2, is that if the guest changes its MAC address on the fly, it merely needs to send a packet and the bridge will learn its new MAC address. That said our firewall rules can potentially block this dynamic change. With the new impl 1. The bridge has a table mapping MAC <-> TAP devices which is initially empty 2. Libvirt tells the bridge what MAC address is associated with each TAP device added 3. The bridge always forwards the packets to the correct TAP device 4. The physical device is never promiscuous, it is programmed with MACs of all TAP devices In both cases the guest can send packets with spoofed MAC addresses. In the old code these packets would make it onto the LAN and it can receive replies back. In the new code these packets would make it onto the LAN but it will never receive replies back. At a high level the conceptual difference is whether the table of MAC addresses & TAP devices is statically defined by libvirt or dynamically defined on the fly. > HERE IS THE REQUEST FOR OPINIONS/IDEAS: > > This V2 of the patchset addresses several issues brought up by jferlan > on the original series, and changes the name of the attribute from: > > promiscLinks='yes|no' > > to > > fdb='learningWithFlood|managed' > > I'm somewhat more happy with this new naming than the previous but > still looking for better ideas. It is closer to describing what the > new code really does, but "learningWithFlood" seems a bit long and > awkward, while I have been told that "fdb" is too short and > unrecognizeable (I will point out that 1) "fdb" is the same name used > by iproute2's "bridge" command, and 2) another bridge option, "stp" is > also a three letter acronym that will only be recognized by those > familiar with configuring an L2 bridge device or watching NASCAR on > Saturday afternoons (or whenever it's on - not a fan myself :-)) One thing to remember is that libvirt is supposed to be providing a higher level configuration language that is independant of specific implementations. 'fdb' is terminology specific to the Linux bridge implementation, so I don't think that's appropriate to use in the XML configuration.

There is a user visible change in behaviour because this new solution, as implemented in this patch series, will no longer allow a guest to change its MAC address on the fly. Of course if QEMU can provide a notification to libvirt when the guest changes its MAC address, then libvirt can update the MAC table and so re-gain functional equivalance with the old solution. I think we would want to be able to turn that on or off though.

This all leads me to suggest that we should use the following syntax that is indepedant of the particular implementation choice, and instead represents the logical behaviour of the feature. macTable=static|dynamic The mode of 'static' means that the MAC address table will always match MACs recorded in libvirt guest XML. This is the new mode that your patches implement The mode of 'dynamic' means that the MAC address table will be able to dynamically update itself as the guest changes MAC addresses. In the near term, this will use the traditional bridge learning mode. In the long term, if we can get MAC change notifications from QEMU, we can switch this over to use the new bridge features.