libvirt users - Jun 2015 - Recommended change for the networking page in wiki

Hello all,

I was told on IRC that I should come here to discuss a recommended change
on the networking page in the wiki.

If you take a look at the "Basic Script" shown here:
http://wiki.libvirt.org/page/Networking#Forwarding_Incoming_Connections

It does two things:
1) Create a DNAT rule in the NAT table of IPTABLES
2) Create a FORWARD rule in the FILTER table of IPTABLES

The FORWARD rule is set up as it ought to be, however, the DNAT rule has
some unintended consequences. I set up a DNAT on port 80, and suddenly, I
couldn't access out on port 80 anymore from my guest machine. However, if I
changed the destination address from "anywhere" to the IP of the host
machine, the problem resolved. So I change the script to as follows.
(Changes are highlighted. For some reason the original script didn't work
using /bin/sh, but it did with /bin/bash, so I changed that too).

*#!/bin/bash*
# used some from advanced script to have multiple ports: use an equal
number of guest and host ports

Guest_name=xxxxxxx
Guest_ipaddr=xxx.xxx.xxx.xx
*Host_ipaddr=xxx.xxx.xxx.xx*
Host_port=( '80' '443' )
Guest_port=( '80' '443' )
length=$(( ${#Host_port[@]} - 1 ))
if [ "${1}" = "${Guest_name}" ]; then
   if [ "${2}" = "stopped" -o "${2}" =
"reconnect" ]; then
       for i in `seq 0 $length`; do
               iptables -t nat -D PREROUTING *-d ${Host_ipaddr}* -p tcp
--dport ${Host_port[$i]} -j DNAT --to ${Guest_ipaddr}:${Guest_port[$i]}
               iptables -D FORWARD -d ${Guest_ipaddr}/32 -p tcp -m state
--state NEW -m tcp --dport ${Guest_port[$i]} -j ACCEPT
       done
   fi
   if [ "${2}" = "start" -o "${2}" =
"reconnect" ]; then
       for i in `seq 0 $length`; do
               iptables -t nat -A PREROUTING *-d ${Host_ipaddr}* -p tcp
--dport ${Host_port[$i]} -j DNAT --to ${Guest_ipaddr}:${Guest_port[$i]}
               iptables -I FORWARD 4 -d ${Guest_ipaddr}/32 -p tcp -m state
--state NEW -m tcp --dport ${Guest_port[$i]} -j ACCEPT
       done
   fi
fi


Lastly, I should note that I am using Ubuntu 14.04, both for the host and
guest.

I'm also curious as to why this is considered a hack method. It states in
the wiki that "This method is a hack", but it doesn't express why.
Many VM
Servers have similar features. I know Virtual Box does, I use the same
feature there. It may not be how I would set up a production server, but
doesn't make it a hack.

Thanks,
BJ

Laine, you make a fair point about the hack. You are right, until this is
natively supported by libvirt, I suppose it is a hack. I guess I've just
grown very accustomed to hacks in linux. :-) That said, it seems like I've
found the right person to look at the alterations I've made to the script.

If you think I should update the wiki page to show the changes I've made to
the script, I'll go ahead and do it. If you'd rather maintain it,
I'm happy
to let you do so. The change does seem logical to me. The original DNAT
rule that it set up on my ubuntu machine basically said: "route port 80
traffic from anywhere to anywhere to the guest machine", which with that in
place prevented me from using apt-get on the guest machine. After I made
the change, everything seemed to work fine (port 80 traffic to the machine
as well as apt-get). Of course, I'm only testing this on Ubuntu, and
that's
probably my biggest blind spot here. I have no idea if other distros might
be slightly different.

Thanks,
BJ

On Mon, Jun 8, 2015 at 2:43 PM, Laine Stump <laine@laine.org> wrote:
>  On 06/08/2015 03:24 PM, BJ wrote:
>
> Hello all,
>
>  I was told on IRC that I should come here to discuss a recommended
> change on the networking page in the wiki.
>
>  If you take a look at the "Basic Script" shown here:
> http://wiki.libvirt.org/page/Networking#Forwarding_Incoming_Connections
>
>  It does two things:
> 1) Create a DNAT rule in the NAT table of IPTABLES
> 2) Create a FORWARD rule in the FILTER table of IPTABLES
>
>  The FORWARD rule is set up as it ought to be, however, the DNAT rule has
> some unintended consequences. I set up a DNAT on port 80, and suddenly, I
> couldn't access out on port 80 anymore from my guest machine.
>
>
> I'm not seeing the same results here. outbound port 80 is still allowed
> from my host once the rule is in place. Of course incoming port 80
isn't
> being redirected to the guest either, so I think I have bigger problems.
> (this is likely due to the fact that I haven't tried the script again
since
> 2013 or so, and a lot has likely changed with the iptables rules added to a
> Fedora system since then)
>
>   However, if I changed the destination address from "anywhere"
to the IP
> of the host machine, the problem resolved. So I change the script to as
> follows. (Changes are highlighted. For some reason the original script
> didn't work using /bin/sh, but it did with /bin/bash, so I changed that
> too).
>
>
> I don't know for sure, but my guess is that this line:
>
>   length=$(( ${#Host_port[@]} - 1 ))
>
> which was added by vgerris in order to support forwarding of multiple
> ports, could be what's causing the incompatibility (that wasn't in
the
> original, simpler version of the script, written by me.)
>
>
>  *#!/bin/bash*
> # used some from advanced script to have multiple ports: use an equal
> number of guest and host ports
>
>  Guest_name=xxxxxxx
> Guest_ipaddr=xxx.xxx.xxx.xx
> *Host_ipaddr=xxx.xxx.xxx.xx*
> Host_port=( '80' '443' )
> Guest_port=( '80' '443' )
> length=$(( ${#Host_port[@]} - 1 ))
> if [ "${1}" = "${Guest_name}" ]; then
>    if [ "${2}" = "stopped" -o "${2}" =
"reconnect" ]; then
>        for i in `seq 0 $length`; do
>                iptables -t nat -D PREROUTING *-d ${Host_ipaddr}* -p tcp
> --dport ${Host_port[$i]} -j DNAT --to ${Guest_ipaddr}:${Guest_port[$i]}
>                iptables -D FORWARD -d ${Guest_ipaddr}/32 -p tcp -m state
> --state NEW -m tcp --dport ${Guest_port[$i]} -j ACCEPT
>        done
>    fi
>    if [ "${2}" = "start" -o "${2}" =
"reconnect" ]; then
>        for i in `seq 0 $length`; do
>                iptables -t nat -A PREROUTING *-d ${Host_ipaddr}* -p tcp
> --dport ${Host_port[$i]} -j DNAT --to ${Guest_ipaddr}:${Guest_port[$i]}
>                iptables -I FORWARD 4 -d ${Guest_ipaddr}/32 -p tcp -m state
> --state NEW -m tcp --dport ${Guest_port[$i]} -j ACCEPT
>        done
>    fi
> fi
>
>
>  Lastly, I should note that I am using Ubuntu 14.04, both for the host
> and guest.
>
>  I'm also curious as to why this is considered a hack method. It states
> in the wiki that "This method is a hack", but it doesn't
express why.
>
>
> I consider it a hack because:
>
> 1) It requires the IP address of the guest to be known before the guest is
> started, so either you need to guess the guest's IP (if the guest is
> getting its IP address via dhcp) or configured the guest IP address in
> multiple places.
>
> 2) It requires using a hook script, which "taints" *all* guests
on this
> machine, rendering them "unsupported" in the eyes of some (in
practice,
> once you see what is causing the tainting it's not such a big deal, but
it
> does mean that an external script gets control with elevated privileges).
>
> 3) The "configuration" is just the contents of the shell script,
so it
> can't be supported by any higher level management application - you
will
> always need to directly modify this single shell script file.
>
> 4) If you change the configuration for a guest while that guest is
> running, any forwarding rules that were a part of the old config but not in
> the new config will be orphaned in your iptables nat table until *all*
> rules are flushed.
>
> 5) If anything else messes with the iptables rules, these port forwarding
> rules are broken. Especially on older versions of libvirt (0.9.12 and
> older, which is around the time that entry was written), just restarting
> libvirtd would break the port forwarding.
>
> 6) As both of us have experienced here, interaction with the iptables
> rules of the underlying system can easily result in it not working at all.
>
>
> The proper way to support this would be XML added to the domain
> configuration, something like:
>
>    <interface type='network'>
>      <source network='default'/>
>      <model type='virtio'/
>      <mac address='52:54:00:11:22:33'/>
>      <redirect protocol='tcp'>
>        <public address='1.2.3.4' port='2200'/>
>        <private port='22'/>
>      </redirect>
>      ...
>   </interface>
>
> (or something like that). libvirt would then auto-determine the IP address
> of the interface and add the necessary iptables rules (or perhaps an rinetd
> config line item, which may be less prone to sabotage by distro-specific
> default iptables setups). The inverse would be automatically done when the
> domain was shutdown.
>
>
>      Many VM Servers have similar features. I know Virtual Box does, I
> use the same feature there. It may not be how I would set up a production
> server, but doesn't make it a hack.
>
>
> It's not the port forwarding itself that I consider to be a hack,
it's the
> method that's used by that script to accomplish it. (And since I was
the
> original author of the script and the entry in the wiki, I think I am
> allowed to criticize it :-)
>

On 06/08/2015 02:43 PM, Laine Stump wrote:
>> However, if I changed the destination address from "anywhere"
to the
>> IP of the host machine, the problem resolved. So I change the script
>> to as follows. (Changes are highlighted. For some reason the original
>> script didn't work using /bin/sh, but it did with /bin/bash, so I
>> changed that too).
> 
> I don't know for sure, but my guess is that this line:
> 
>   length=$(( ${#Host_port[@]} - 1 ))
Correct - that line is a bashism, and is not portable when /bin/sh is dash.
> 
> which was added by vgerris in order to support forwarding of multiple
> ports, could be what's causing the incompatibility (that wasn't in
the
> original, simpler version of the script, written by me.)
> 
>>
>> *#!/bin/bash*
>> # used some from advanced script to have multiple ports: use an equal
>> number of guest and host ports
>>
>> Guest_name=xxxxxxx
>> Guest_ipaddr=xxx.xxx.xxx.xx
>> *Host_ipaddr=xxx.xxx.xxx.xx*
>> Host_port=( '80' '443' )
>> Guest_port=( '80' '443' )
In fact, these two lines are also bashisms.  All the more reason to
require bash.
>> length=$(( ${#Host_port[@]} - 1 ))
>> if [ "${1}" = "${Guest_name}" ]; then
>>    if [ "${2}" = "stopped" -o "${2}" =
"reconnect" ]; then
test ... -o ... (also spelled [ ... -o ... ]) is not portable, even on
bash.  There are some expressions that are completely ambiguous on how
to be parsed, when -o is in the mix.  It is better to spell it: [ ... ]
|| [ ... ] (that is, use the shell's || instead of test's -o to do the
conjunction).
>>        for i in `seq 0 $length`; do
seq is not portable outside of GNU/Linux systems.
>> Lastly, I should note that I am using Ubuntu 14.04, both for the host
>> and guest. 
Yep, that's a system that uses dash for /bin/sh.
>>
>> I'm also curious as to why this is considered a hack method. It
states
>> in the wiki that "This method is a hack", but it doesn't
express why.
> 
> I consider it a hack because:
> 
> 1) It requires the IP address of the guest to be known before the guest
> is started, so either you need to guess the guest's IP (if the guest is
> getting its IP address via dhcp) or configured the guest IP address in
> multiple places.
Although recent work has been made to get libvirt to add an API that
queries the guest for its IP address, once the guest is running.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org