How IPTables Work with libVirt

Posted:   |  More posts about Snippets

I saw this on a mailing list, and I thought it'd be great to keep and share.

[libvirt] FYI: a short guide to libvirt & network filtering iptables/ebtables use

    * From: "Daniel P. Berrange" <berrange redhat com>
    * To: libvir-list redhat com
    * Subject: [libvirt] FYI: a short guide to libvirt & network filtering iptables/ebtables use
    * Date: Wed, 30 Jun 2010 16:04:47 +0100

I just wrote this to assist some Red Hat folks understanding
what libvirt does with iptables, and thought it is useful info
for the whole libvirt community. When I have time I'll adjust
this content so that it can fit into the website in relevant

        Firewall / network filtering in libvirt

There are three pieces of libvirt functionality which do network
filtering of some type. At a high level they are:

 - The virtual network driver.

   This provides a isolated bridge device (ie no physical NICs
   enslaved). Guest TAP devices are attached to this bridge.
   Guests can talk to each other & the host, and optionally the
   wider world.

 - The QEMU driver MAC filtering

   This provides a generic filtering of MAC addresses to prevent
   the guest spoofing its MAC address. This is mostly obsoleted by
   the next item, so won't be discussed further.

 - The network filter driver

   This provides fully configurable, arbitrary network filtering
   of traffic on guest NICs. Generic rulesets are defined at the
   host level to control traffic in some manner. Rules sets are
   then associated with individual NICs of a guest. While not as
   expressive as directly using iptables/ebtables, this can still
   do nearly everything you would want to on a guest NIC filter.

The virtual network driver

The typical configuration for guests is to use bridging of the
physical NIC on the host to connect the guest directly to the LAN.
In RHEL6 there is also the possibility of using macvtap/sr-iov
and VEPA connectivity. None of this stuff plays nicely with wireless
NICs, since they will typically silently drop any traffic with a
MAC address that doesn't match that of the physical NIC.

Thus the virtual network driver in libvirt was invented. This takes
the form of an isolated bridge device (ie one with no physical NICs
enslaved). The TAP devices associated with the guest NICs are attached
to the bridge device. This immediately allows guests on a single host
to talk to each other and to the host OS (modulo host IPtables rules).

libvirt then uses iptables to control what further connectivity is
available. There are three configurations possible for a virtual
network at time of writing

  - isolated: all off-node traffic is completely blocked
  - nat: outbound traffic to the LAN is allowed, but MASQUERADED
  - forward: outbound traffic to the LAN is allowed

The latter 'forward' case requires the virtual network be on a
separate sub-net from the main LAN, and that the LAN admin has
configured routing for this subnet.  In the future we intend to
add support for IP subnetting and/or proxy-arp. This allows for
the virtual network to use the same subnet as the main LAN &
should avoid need for the LAN admin to configure special routing.

Libvirt will optionally also provide DHCP services to the virtual
network using DNSMASQ. In all cases, we need to allow DNS/DHCP
queries to the host OS. Since we can't predict whether the host
firewall setup is already allowing this, we insert 4 rules into
the head of the INPUT chain

  target     prot opt in     out     source               destination
  ACCEPT     udp  --  virbr0 *             udp dpt:53
  ACCEPT     tcp  --  virbr0 *             tcp dpt:53
  ACCEPT     udp  --  virbr0 *             udp dpt:67
  ACCEPT     tcp  --  virbr0 *             tcp dpt:67

Note we have restricted our rules to just the bridge associated
with the virutal network, to avoid opening undesirable holes in
the host firewall wrt the LAN/WAN.

The next rules depend on the type of connectivity allowed, and go
in the main FORWARD chain:


Allow traffic between guests. Deny inbound. Deny outbound.

  target     prot opt in     out     source               destination
  ACCEPT     all  --  virbr1 virbr1  
  REJECT     all  --  *      virbr1             reject-with icmp-port-unreachable
  REJECT     all  --  virbr1 *             reject-with icmp-port-unreachable


Allow inbound related to an established connection. Allow
outbound, but only from our expected subnet. Allow traffic
between guests. Deny all other inbound. Deny all other outbound.

  target     prot opt in     out     source               destination
  ACCEPT     all  --  *      virbr0      state RELATED,ESTABLISHED
  ACCEPT     all  --  virbr0 *
  ACCEPT     all  --  virbr0 virbr0  
  REJECT     all  --  *      virbr0             reject-with icmp-port-unreachable
  REJECT     all  --  virbr0 *             reject-with icmp-port-unreachable


Allow inbound, but only to our expected subnet. Allow
outbound, but only from our expected subnet. Allow traffic
between guests. Deny all other inbound. Deny all other outbound.

  target     prot opt in     out     source               destination
  ACCEPT     all  --  *      virbr2  
  ACCEPT     all  --  virbr2 *
  ACCEPT     all  --  virbr2 virbr2  
  REJECT     all  --  *      virbr2             reject-with icmp-port-unreachable
  REJECT     all  --  virbr2 *             reject-with icmp-port-unreachable

Finally, with type=nat, there is also an entry in the POSTROUTING
chain to apply masquerading

  target     prot opt in     out     source               destination
  MASQUERADE all  --  *      *   &nbsp;!

The network filter driver

This driver provides a fully configurable network filtering capability
that leverages ebtables, iptables and ip6tables. This was written by
the libvirt guys at IBM and although its XML schema is defined by libvirt,
the conceptual model is closely aligned with the DMTF CIM schema for
network filtering

The filters are managed in libvirt as a top level, standalone object.
This allows the filters to then be referenced by any libvirt object
that requires their functionality, instead tieing them only to use
by guest NICs. In the current implementation, filters can be associated
with individual guest NICs via the libvirt domain XML format. In the
future we might allow filters to be associated with the virtual network
objects. Further we're expecting to define a new 'virtual switch' object
to remove the complexity of configuring bridge/sriov/vepa networking
modes. This make also end up making use of network filters.

There are a new set of virsh commands for managing network filters

   virsh nwfilter-define     define or update a network filter from an XML file
   virsh nwfilter-undefine   undefine a network filter
   virsh nwfilter-dumpxml    network filter information in XML
   virsh nwfilter-list       list network filters
   virsh nwfilter-edit       edit XML configuration for a network filter

There are equivalently named C APIs for each of these commands.

As with all objects libvirt manages, network filters are configured
using an XML format. At a high level the format looks like this:

  <filter name='no-spamming' chain='XXXX'>


    <rule ...>

    <filterref filter='XXXX'/>


Every filter has a name and UUID which serve as unique identifiers.
A filter can have zero-or-more <rule> elements which are used to
actually define network controls. Filters can be arranged into a
DAG, so zero-or-more <filterref/> elements are also allowed. Cycles
in the graph are not allowed.

The <rule> element is where all the interesting stuff happens. It
has three attributes, an action, a traffic direction and an optional
priority. eg

  <rule action='drop' direction='out' priority='500'>

Within the rule there are a wide variety of elements allowed, which
do protocol specific matching. Supported protocols currently include
'mac', 'arp', 'rarp', 'ip', 'ipv6', 'tcp/ip', 'icmp/ip', 'igmp/ip',
'udp/ip', 'udplite/ip' 'esp/ip', 'ah/ip', 'sctp/ip', 'tcp/ipv6',
'icmp/ipv6', 'igmp/ipv6', 'udp/ipv6', 'udplite/ipv6', 'esp/ipv6',
'ah/ipv6', 'sctp/ipv6'. Each protocol defines what is valid inside
the <rule> element, the general pattern though is

   <protocol match='yes|no' attribute1='value1' attribute2='value2'/>

so, eg a TCP protocol, matching ports 0-1023 would be expressed

   <tcp match='yes' srcportstart='0' srcportend='1023'/>

Attributes can included references to variables defined by the
object using the rule. So the guest XML format allows each NIC
to have a MAC address and IP address defined. These are made
available to filters via the variables $IP and $MAC.

So to define a filter the prevents IP address spoofing we can
simply match on source IP address&nbsp;!= $IP

  <filter name='no-ip-spoofing' chain='ipv4'>
    <rule action='drop' direction='out'>

        <ip match='no' srcipaddr='$IP' />

I'm not going to go into details on all the other protocol
matches you can do, because it'll take far too much space.
You can read about the options here

Out of the box in RHEL6/Fedora rawhide, libvirt ships with a
set of default useful rules

  # virsh nwfilter-list
  UUID                                  Name
  15b1ab2b-b1ac-1be2-ed49-2042caba4abb  allow-arp
  6c51a466-8d14-6d11-46b0-68b1a883d00f  allow-dhcp
  7517ad6c-bd90-37c8-26c9-4eabcb69848d  allow-dhcp-server
  3d38b406-7cf0-8335-f5ff-4b9add35f288  allow-incoming-ipv4
  5ff06320-9228-2899-3db0-e32554933415  allow-ipv4
  db0b1767-d62b-269b-ea96-0cc8b451144e  clean-traffic
  f88f1932-debf-4aa1-9fbe-f10d3aa4bc95  no-arp-spoofing
  772f112d-52e4-700c-0250-e178a3d91a7a  no-ip-multicast
  7ee20370-8106-765d-f7ff-8a60d5aaf30b  no-ip-spoofing
  d5d3c490-c2eb-68b1-24fc-3ee362fc8af3  no-mac-broadcast
  fb57c546-76dc-a372-513f-e8179011b48a  no-mac-spoofing
  dba10ea7-446d-76de-346f-335bd99c1d05  no-other-l2-traffic
  f5c78134-9da4-0c60-a9f0-fb37bc21ac1f  no-other-rarp-traffic
  7637e405-4ccf-42ac-5b41-14f8d03d8cf3  qemu-announce-self
  9aed52e7-f0f3-343e-fe5c-7dcb27b594e5  qemu-announce-self-rarp

Most of these are just building blocks. The interesting one here
is 'clean-traffic'. This pulls together all the building blocks
into one filter that you can then associate with a guest NIC.
This stops the most common bad things a guest might try, IP
spoofing, arp spoofing and MAC spoofing. To look at the rules for
any of these just do

   virsh nwfilter-dumpxml FILTERNAME|UUID

They are all stored in /etc/libvirt/nwfilter, but don't edit
files there directly. Use 'virsh nwfilter-define' to update
them. This ensures the guests have their iptables/ebtables
rules recreated.

To associate the clean-trafffic filter with a guest, edit the
guest XML config and change the <interface> element to include
a <filterref> and also specify the whitelisted <ip addres/> the
guest is allowed to use

    <interface type='bridge'>

      <mac address='52:54:00:56:44:32'/>
      <source bridge='br1'/>
      <ip address=''/>
      <target dev='vnet0'/>
      <model type='virtio'/>
      <filterref filter='clean-traffic'/>


If no <ip address> is included, the network filter driver will
activate its 'learning mode'. This uses libpcap to snoop on
network traffic the guest sends and attempts to identify the
first IP address it uses. It then locks traffic to this address.
Obviously this isn't entirely secure, but it does offer some
protection against the guest being trojaned once up & running.
In the future we intend to enhance the learning mode so that it
looks for DHCPOFFERS from a trusted DHCP server and only allows
the offered IP address to be used.

Now, how is all this implemented... The network filter driver
uses a combination of ebtables, iptables and ip6tables, depending
on which protocols are referenced in a filter. The out of the box
'clean-traffic' filter rules only require use of ebtables. If you
want to do matching at tcp/udp/etc protocols (eg to add a new
filter 'no-email-spamming' to block port 25), then iptables will
also be used.

The driver attempts to keep its rules separate from those that
the host admin might already have configured. So the first thing
it does with ebtables, is to add two hooks in POSTROUTING &
PREROUTING chains, to redirect traffic to custom chains. These
hooks match on the TAP device name of the guest NIC, so they
should not interact badly with any administrator defined rules

  Bridge chain: PREROUTING, entries: 1, policy: ACCEPT
  -i vnet0 -j libvirt-I-vnet0

  Bridge chain: POSTROUTING, entries: 1, policy: ACCEPT
  -o vnet0 -j libvirt-O-vnet0

To keep things managable & easy to follow, the driver will then
create further sub-chains for each protocol then it needs to match

  Bridge chain: libvirt-I-vnet0, entries: 5, policy: ACCEPT
  -p IPv4 -j I-vnet0-ipv4
  -p ARP -j I-vnet0-arp
  -p 0x8035 -j I-vnet0-rarp
  -p 0x835 -j ACCEPT
  -j DROP

  Bridge chain: libvirt-O-vnet0, entries: 4, policy: ACCEPT
  -p IPv4 -j O-vnet0-ipv4
  -p ARP -j O-vnet0-arp
  -p 0x8035 -j O-vnet0-rarp
  -j DROP

Finally, come the actual implementation of the filters. These
example is showing the 'clean-traffic' filter implementation.
I'm not going to explain what this is doing now&nbsp;:-)

  Bridge chain: I-vnet0-ipv4, entries: 2, policy: ACCEPT
  -s&nbsp;! 52:54:0:56:44:32 -j DROP
  -p IPv4 --ip-src&nbsp;! -j DROP

  Bridge chain: O-vnet0-ipv4, entries: 1, policy: ACCEPT

  Bridge chain: I-vnet0-arp, entries: 6, policy: ACCEPT
  -s&nbsp;! 52:54:0:56:44:32 -j DROP
  -p ARP --arp-mac-src&nbsp;! 52:54:0:56:44:32 -j DROP
  -p ARP --arp-ip-src&nbsp;! -j DROP
  -p ARP --arp-op Request -j ACCEPT
  -p ARP --arp-op Reply -j ACCEPT
  -j DROP

  Bridge chain: O-vnet0-arp, entries: 5, policy: ACCEPT
  -p ARP --arp-op Reply --arp-mac-dst&nbsp;! 52:54:0:56:44:32 -j DROP
  -p ARP --arp-ip-dst&nbsp;! -j DROP
  -p ARP --arp-op Request -j ACCEPT
  -p ARP --arp-op Reply -j ACCEPT
  -j DROP

  Bridge chain: I-vnet0-rarp, entries: 2, policy: ACCEPT
  -p 0x8035 -s 52:54:0:56:44:32 -d Broadcast --arp-op Request_Reverse --arp-ip-src --arp-ip-dst --arp-mac-src 52:54:0:56:44:32 --arp-mac-dst 52:54:0:56:44:32 -j ACCEPT
  -j DROP

  Bridge chain: O-vnet0-rarp, entries: 2, policy: ACCEPT
  -p 0x8035 -d Broadcast --arp-op Request_Reverse --arp-ip-src --arp-ip-dst --arp-mac-src 52:54:0:56:44:32 --arp-mac-dst 52:54:0:56:44:32 -j ACCEPT
  -j DROP

NB, we would have liked to include the prefix 'libvirt-' in all
of our chain names, but unfortunately the kernel limits names
to a very short maximum length. So only the first two custom
chains can include that prefix. The others just include the
TAP device name + protocol name.

If I define a new filter 'no-spamming' and then add this to the
'clean-traffic' filter, I can illustrate how iptables usage works.

  # cat > /root/spamming.xml <<EOF
  <filter name='no-spamming' chain='root'>

    <rule action='drop' direction='out' priority='500'>
      <tcp dstportstart='25' dstportend='25'/>

  # virsh nwfilter-define /root/spamming.xml
  # virsh nwfilter-edit clean-traffic

   ...add  <filterref filter='no-spamming'/>

All active guests immediately have their iptables/ebtables rules

The network filter driver deals with iptables in a very similar
way. First it separates out its rules from those the admin may
have defined, by adding a couple of hooks into the INPUT/FORWARD

  Chain INPUT (policy ACCEPT 13M packets, 21G bytes)
  target           prot opt in     out     source               destination
  libvirt-host-in  all  --  *      *  

  Chain FORWARD (policy ACCEPT 5532K packets, 3010M bytes)
  target           prot opt in     out     source               destination
  libvirt-in       all  --  *      *  
  libvirt-out      all  --  *      *  
  libvirt-in-post  all  --  *      *  

These custom chains, then do matching based on the TAP device
name, so they won't open holes in the admin defined matches for
the LAN/WAN (if any).

Chain libvirt-host-in (1 references)
  target     prot opt in     out     source               destination
  HI-vnet0   all  --  *      *             [goto] PHYSDEV match --physdev-in vnet0

Chain libvirt-in (1 references)
  target     prot opt in     out     source               destination
  FI-vnet0   all  --  *      *             [goto] PHYSDEV match --physdev-in vnet0

Chain libvirt-in-post (1 references)
  target     prot opt in     out     source               destination
  ACCEPT     all  --  *      *             PHYSDEV match --physdev-in vnet0

Chain libvirt-out (1 references)
  target     prot opt in     out     source               destination
  FO-vnet0   all  --  *      *             [goto] PHYSDEV match --physdev-out vnet0

Finally, we can see the interesting bit which is the actual
implementation of my filter to block port 25 access:

Chain FI-vnet0 (1 references)
  target     prot opt in     out     source               destination
  DROP       tcp  --  *      *             tcp dpt:25

Chain FO-vnet0 (1 references)
  target     prot opt in     out     source               destination
  DROP       tcp  --  *      *             tcp spt:25

Chain HI-vnet0 (1 references)
  target     prot opt in     out     source               destination
  DROP       tcp  --  *      *             tcp dpt:25

One thing in looking at this that you may notice is that if there
are many guests all using the same filters, we will be duplicating
the iptables rules over & over for each guest. This is merely a
limitation of the current rules engine implementation. At the libvirt
object modelling level you can clearly see we've designed the model
so that filter rules are define in one place, and indirectly referenced
by guests. Thus it should be possible to change the impl in the future
so that we can share the actual iptables/ebtables rules for each
guest to create a more scalable system. The stuff in current libvirt
is more or less the very first working impl we've had of this stuff,
so there's not been much optimization work yet.

Also notice that at the XML level we don't expose the fact that we
are using iptables or ebtables at all. The rule definition is done in
terms of network protocols. Thus if we ever find a need, we could
plug in an alternative implementation that calls out to a different
firewall implementation instead of ebtables/iptables (providing that
impl was suitably expressive of course)

Finally, in terms of problems we have in deployment. The biggest
problem is that if the admin does 'service iptables restart' all
our work gets blown away. We've experimented with using lokkit
to record our custom rules in a persistent config file, but that
caused different problem. Admins who were not using lokkit for
their config found that all their own rules got blown away. So
we threw away our lokkit code. Instead we document that if you
run 'service iptables restart', you need to send SIGHUP to libvirt
to make it recreate its rules.

Finally a reminder, that the main documentation we have on this
is online at

|: Red Hat, Engineering, London    -o-;:|
|: -o- -o-;:|
|:        -o-;:|
|: GnuPG: 7D3B9505  -o-   F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505&nbsp;:|o as you would guess this utility also has options to automatically check the validity of tables and repair them.
Comments powered by Disqus
Contents © 2013 Daniel Devine - Nikola Powered - Flattr Me! Flattr this Source