11:03 a.m.
On Mon, Jan 11, 2010 at 12:55:27PM -0500, Stefan Berger wrote:
Hello!
The following is a proposal to introduce network traffic filtering
capabilities for the network traffic originating from and destined to
virtual machines. Libvirt's capabilities will be extended to allow users
to describe what traffic filtering rules are to be applied on a virtual
machine (using XML) and libvirt then creates the appropriate ebtables and
iptables rules and applies those on the host when the virtual machine
starts up, resumes after suspension or resumes on a new host after
migration. libvirt tears down the traffic filtering rules when the virtual
machine shuts down, suspends, or a VM is migrated to another host. It will
also be possible to modify the filtering rules while a virtual machine is
running. In this architecture we apply the firewall rules on the outside
of the virtual machines on the Linux host and make use of the fact that
virtual machines can be configured by libvirt to have their network
traffic pass through the host and the host exposes (tap) 'backend'
interfaces on which the firewall rules can be applied. The application of
the firewall rules is optional and those who do not want to introduce a
raw network throughput performance hit on their system due to the
evaluation of every packet passing through the filtering chains, do not
have to use these capabilities. An initial implementation would be done
for libvirt's Qemu support.
It's relatively clear that this would not work with devices exposed
natively to the domain, say though PCI passthough. I'm wondering what
other case of limitiations could be found. Also this may not map well
for other kind of hypervisors like VMWare, right ?
As stated above, the application of firewall rules on virtual
machines
will require some form of XML description that lets the user choose what
type of filtering is to be performed. In effect the user needs to be able
to tell libvirt which ebtables and iptables rules to generate so that the
appropriate filtering can be done. We realize that ebtables and iptables
have lots of capabilities but think that we need to restrict which
capabilities can actually be 'reached' through this XML.
The following proposal is based on an XML as defined by the DMTF (
http://www.dmtf.org/standards/cim/cim_schema_v2230/CIM_Network.pdf, slide
10) with extensions for processing of ARP packets.It gives control over
the evaluation of Ethernet (MAC) frames, ARP packet data and IP header
information. A (draft) XML skeleton in this case could look as follows:
Humpf, extending astandard XML schemas is really bad practice. Even if
you think that the extension may be accepted later, it may have
different semantic for the same construct and that's a disaster.
I would at least put the new elements in a separate namespace to
avoid something insane if DMTF extends its specification.
<FilterEntry>
<Action>DROP|ACCEPT</Action> <!-- from FilterEntry -->
<TrafficType>incoming [to VM]|outgoing [from VM]</TrafficType>
<Hdr8021Filter>
<HdrSrcMACAddr8021> </HdrSrcMACAddr8021>
<HdrSrcMACMask8021> </HdrSrcMACMask8021>
<HdrDestMACAddr8021> </HdrDestMACAddr8021>
<HdrDestMACMask8021> </HdrDestMACMask8021>
<HdrProtocolID8021> numeric and/or string?
</HdrProtocolID8021>
<HdrPriorityValue8021></HdrPriorityValue8021>
<HdrVLANID8021> </HdrVLANID8021>
</Hdr8021Filter>
<ARPFilter>
<HWType> </HWType>
<ProtocolType> </ProtocolType>
<OPCode> </OPCode>
<SrcMACAddr> </SrcMACAddr>
<SrcIPAddr> </SrcIPAddr>
<DestMACAddr> </DestMACAddr>
<DestIPAddr> </DestIPAddr>
</ARPFilter>
<IPHeadersFilter>
<HdrIPVersion> </HdrIPVersion>
<HdrSrcAddress> </HdrSrcAddress>
<HdrSrcMask> </HdrSrcMask>
<HdrDestAddress> </HdrDestAddress>
<HdrDestMask> </HdrDestMask>
<HdrProtocolID> numeric and/or string? </HdrProtocolID>
<HdrSrcPortStart> </HdrSrcPortStart>
<HdrSrcPortEnd> </HdrSrcPortEnd>
<HdrDestPortStart> </HdrDestPortStart>
<HdrDestPortEnd> </HdrDestPortEnd>
<HdrDSCP> </HdrDSCP>
</IPHeadersFilter>
</FilterEntry>
hum the types for each values should be made explicit if possible,
But in general I find that extremely bulky for a single entry, while
I would expect any domain to have a dozen filters or so just for
simple stuff. I wonder if empty default should just be dropped, so
that basically if you're filtering on a MAC address you don't carry
a bunch of unused fields.
Since the ebtables and iptables command cannot accept all possible
parameters at the same time, only a certain subset of the above parameters
may be set for any given filter command. Higher level application writers
will likely use a library that lets them choose which features they would
want to have enforced, such as no-broadcast or no-multicast, enforcement
of MAC spoofing prevention or ARP poisoning prevention, which then
generates this lower level XML rules in the appropriate order of the
rules.
I wonder if a tailored set of simpler XML matching what is actually
possible and getting away from the DMTF records wouldn't be better,
since the original spec is extended on the feature side and the
applicability is restricted on the implementation side, I'm not sure I
see the point of trying to be compatible.
Further, we will introduce filter pools where a collection of the
above
filter rules can be stored and referenced to by virtual machines'
individual interface. A reference to such a filter pool entry will be
given in the interface description and may look as in the following draft
XML:
<interface type='bridge'>
<source bridge='virbr0'/>
<script path='vif-bridge'/>
<firewall>
<reference profile='generic-layer2' ip_address='10.0.0.1'/>
<reference profile='VM-specific-layer3'/>
</firewall>
</interface>
the devil is in the details, basically how you would template
the profiles based on the strings in the reference. I.e. what the
profile would look like and how for example ip_address='10.0.0.1'
would be instancied in the profile use.
The above XML has one reference to a generic layer2 filter template
XML
that should be usable by multiple virtual machines but would need to be
customized with interface-specific parameters. In this case, the IP
address of the interface is explicitly provided in order to make the
filter XML template a concrete XML for the particular interface. The MAC
address of the interface is not explicitly provided since it will already
have been randomly generated by libvirt at the point when this layer2
filter XML needs to be converted into concrete ebtables commands/rules. We
still need to determine how the format of a template should look like,
though an idea would be to indicate a placeholder for a VM's MAC address
using THIS_MAC and similarly for the IP address with THIS_IP as
placeholder.
Further, we would introduce the management of filter 'pools'. Considering
existing functionality in libvirt and CLI commands for similar type of
management functionality, the following CLI commands would be added:
virsh nwfilter-create <file>
virsh nwfilter-destroy <profile name>
virsh nwfilter-list <options>
virsh nwfilter-dumpxml <profile name>
virsh nwfilter-update <filename> (performs an update on an existing
profile replacing all rules)
possibly also:
virsh nwfilter-edit <profile name>
virsh nwfilter-create-from <profile name>
Please let us know what you think of this proposal.
The feature looks interesting ! It looks it should be applicable to
at least qemu and xen, I'm not so sure about LXC or VirtualBox, and
looks unlikely for VMWare unless they have a matching capability (might
be possible since it's based at least partly on DMTF).
I'm not sure reusing the DMTF format actually helps much, it also has
its risks due to the extension.
Daniel
--
Daniel Veillard | libxml Gnome XML XSLT toolkit http://xmlsoft.org/
daniel(a)veillard.com | Rpmfind RPM search engine http://rpmfind.net/
http://veillard.com/ | virtualization library http://libvirt.org/