[libvirt] libvirt/dnsmasq integration.

As the principal maintainer of dnsmasq, I'm seeing increasing reports of problems on systems which run both dnsmasq and libvirt. I'm fairly sure I understand what's going on in these cases, and I have a few proposals for changes in libvir and dnsmasq that should fix things. The problem is that libvirt runs a private instance of dnsmasq: on machines which are also running a "system" dnsmasq daemon, this can cause problems. Some background: dnsmasq can run in two modes. Default mode: dnsmasq binds the wildcard address and does network magic to determine which interface request packets actually come from, so that the results can be sent back with the correct source address. This has the advantage that network interfaces can come and go and change IP address and dnsmasq will keep working. It's possible to restrict dnsmasq to only reply to requests on some interfaces; requests from other interfaces will be read by dnsmasq and then silently dropped. Telling dnsmasq to use an interface which doesn't exist but might in the future will result in a logged warning, but dnsmasq will still start and when the interface comes up it will work. Bind-interfaces mode: This is the traditional way to do UDP servers. At startup dnsmasq enumerates all the extant interfaces and then opens a socket for each one, listening on the interfaces's IP address. Interfaces may be skipped if excluded by the --interface and --except-interface flags, and any interface specified in --interface which doesn't exist at start-up will generate a fatal error. In almost all cases, default mode is better: --bind-interfaces is only there to cope with old platforms which don't support enough socket options to do default mode. The only time when --bind-interfaces works better is when it's desirable to run more than one instance of dnsmasq or have dnsmasq co-exist with another DNS server. This is not possible in default mode, but it does work in bind-interfaces mode, providing than _all_ instances of dnsmasq are in bind-interfaces mode, and that they listen on a disjoint set of interfaces. Therefore, to allow multiple dnsmasq instances libvirt's private dnsmasq instance is started in bind-interfaces mode: that forces one of the dnsmasq instances to do bind-interfaces. Many of the Linux distibution dnsmasq packages have now implemented an /etc/dnsmasq.d directory where configuration fragments can be dropped. Their libvirt packages are putting a file there which contains a bind-interfaces command, so that the "system" dnsmasq is automatically forced into the same mode, and the two can co-exist. This works, sort-of, but there some disadvantages. Installing libvirt drops the configuration change for the system dnsmasq, but the packages frequently don't restart the system daemon, so that things transiently fail until everything has rebooted. Much worse, the system dnsmasq is forced into bind-interfaces mode and then service to transient interfaces (usb, ad-hoc wifi) no longer works, or, because those interfaces are mentioned in the dnsmasq configuration, dnsmasq now fails at start-up when the interfaces don't exist. I don't think there is a solution to this problem wholly within dnsmasq: the nature of the BSD-sockets universe means that there never be two instances of dnsmasq without bind-interfaces, and it's, at best, difficult to get round the limitations of that WRT transient network interfaces. There is, an alternative solution, which would involve changes to both dnsmasq, libvirt and their packaging. Hence this mail. My proposal is to get rid of the necessity for two dnsmasq instances. Libvirt should check for the existance of a "system" dnsmasq and, if the system daemon exists, libvirt should drop the required configuration into /etc/dnsmasq.d and then restart it. If the system daemon is not installed or enabled, libvirt can start a private instance as now. The difficulty with this scheme is that libvirt needs to create some configuration which enables the services it needs on the virtual network without disturbing, or being disturbed by, whatever configuration exists for the system daemon. That's not currently possible, but it can be made possible. I'm assuming that libvirt needs to provide a set of IP address / MAC address mappings, and range of IP addresses on a virtual network. It needs DHCP and DNS service on the virtual network. The dhcp-host IP/MAC mappings are a non-problem: they will be ignored for any other subnet where the IP addresses don't fit, and any other dhcp-hosts in the system configuration will be similarly ignored for DHCP on the virtual network subnet. The dhcp-range is more of a problem. Service to particular networks in dnsmasq is controlled by interface=<interface name"> lines in the configuration. If there are none of these, service is provided to all interfaces. If they exist, service is limited to the interfaces specified. The existence of any dhcp-range line in dnsmasq's configuration enables the DHCP server for any subnet unless explicitly limited to particular interfaces. So a default dnsmasq installation, (with no interface=<interface>) which provides DNS everywhere but DHCP nowhere would be turned into one which provided DHCP on every interface by libvirt adding a dhcp-range. Since there wouldn't be a suitable DHCP range for most subnets, this would only result in logged errors, but it is still not good. Worse, there's no good answer to the question 'should libvirt include interface=virt0"' in the configuration it supplies? If it does, then the "enable DHCP on all interfaces" problem is solved, but a default system configuration with no interface declaration is transformed from one which provides DNS everywhere to one which provides DNS only to the virtual interface. If libvirt doesn't provide "interface=virt0" and the system configuration includes interface declarations, then there will be no DNS or DHCP service to the virtual network. To solve this, I propose to add an optional interface name to the dhcp-range declaration. The semantics of this would be rather odd, but solve the problem perfectly. 1) for DHCP, if any other dhcp-range exists _without_ an interface name, them the interface name is ignored and and things behave as before, otherwise DHCP is only provided to interfaces mentioned in dhcp-range declarations. 2) for DNS, if there are no interface declarations, things work as before. If there are interface declarations, the interfaces mentioned in dhcp-ranges are added to the set which get DNS service. With these rules, it should be possible for libvirt to drop eg dhcp-range=interface:virt0,192.168.0.1,192.168.0.240 into the configuration of the system dnsmasq and get DHCP and DNS service for virt0, irrespective of any other configuration in the system dnsmasq, and doing so shouldn't affect the services supplied elsewhere. The code in libvirt to make this work looks like this: echo dhcp-range=interface:virt0,<ip range> >>/etc/dnsmasq.d/libvirt if <system dnsmasq is not installed or not enabled> dnsmasq --interface=virt0\ --bind-interfaces --conf-file=/etc/dnsmasq.d/libvirt else /etc/init.d/dnsmasq restart (The --bind-interfaces in the private-dnsmasq instance keeps dnsmasq from clashing with other nameservers eg BIND which may be running.) The system dnsmasq package has to ensure that /etc/dnsmasq.d is read for configuration fragments, and the dnsmasq package and the libvirt package will have to co-operate to manage transitions between private and system dnsmasq mode caused by package installation or removal. Does that make sense? It's a long and involved explanation to come to cold. I fear I may have over-simplified what libvirt is doing with dnsmasq, in which case please enlighten me and I'll modify my scheme to take that into account. If this looks good I can easily have the necessary dnsmasq changes in the next release. Cheers, Simon.