On 06/20/2017 05:27 AM, Martin Kletzander wrote:
> On Tue, Jun 20, 2017 at 10:05:19AM +0200, Martin Kletzander wrote:
>> On Tue, Jun 20, 2017 at 02:26:59AM -0400, Travis S. Johnson wrote:
>>> Hello,
>>>
>>> I came across an interesting problem in my home lab a few weeks ago
>>> as I'm
>>> prepping for my RHCE exam using Michael Jang study guide. I've been
>>> at this
>>> for days now, and I still can't wrap my head around how two or more
>>> virtual
>>> networks in default NAT configuration are even allowed to communicate
>>> with
>>> each other despite what the libvirt documentation said.
>>>
>>>
>>> Here's the excerpt I'm referring to in the wiki link here:
>>> http://wiki.libvirt.org/page/Networking#Forwarding_Incoming_Connections:
>>>
>>>> By default, guests that are connected via a virtual network with
>>>> <forward
>>>> mode='nat'/> can make any outgoing network connection they like.
>>>> Incoming
>>>> connections are allowed from the host, and from other guests
>>>> connected to
>>>> the same libvirt network, but all other incoming connections are
>>>> blocked by
>>>> iptables rules.
>>>
>
> I did not read this properly...
>
>>>
>>> Also here's another assertion from 'The virtual network driver'
>>> section in
>>> http://libvirt.org/firewall.html:
>>>
>>>> type=nat
>>>>
>>>> Allow inbound related to an established connection. Allow outbound, but
>>>> only from our expected subnet. Allow traffic between guests. Deny
>>>> all other
>
> I would expect the 'traffic between guests' to mean only guests on the
> same network.  Also that lines up with what's written above.

That is correct.

>
>>>> inbound. Deny all other outbound.
>>>
>
> [...]
>
>>
>> Thanks for reporting this.  It's clearly a bug in libvirt.  The rules
>> are in this order:
>>
>>  all rules for virbr0
>>  all rules for virbr1
>>  all rules for virbr2
>>
>> But what we should do instead is:
>>
>>  input rules for all networks
>>  local rules for all networks
>>  output rules for all networks
>>  reject rules for all networks

That doesn't achieve the desired results (as you figured out later,
according to your followup comments.


>>
>> The problem is that we do not know how other rules look like.  So what
>> we might need to do is create chains where rules for the first network
>> are, then only append network rules into those chains.
>>
>> Would you mind filing a bug for this issue, so we can properly track it
>> and don't forget about it?  I'll have a look at it in the meantime, but
>> don't promise anything since I'm not that familiar with that part of the
>> codebase.
>>
>
> Actually, now that I'm thinking about it, the problem is not that we
> disallow the communication from virbr1 to virbr0, but the problem is
> that we allow the connection from virbr0 to everywhere.

Yep!

That's actually a known bug that has been there "forever". It has always
been brought up in the context of "I want communication between my NATed
networks, and it only works in one direction!", rather than "I want to
block communication between my NATed networks, but it gets past in one
direction!". When people learn that it's (kind of) blocked, they work
around it by adding a new network that is shared by all the guests.
Because all of the reports have been in this direction, fixing it
properly has been very low on (my) priority list.


>
> Maybe the solution would be: for each network, insert the rules on top
> of the forward chain and for each started network, explicitly reject
> that one.  So that after one network starts it would look like this (I'm
> writing this from memory just to illustrate the idea, not actually
> looking up how stuff looks):
>
>  ACCEPT any virbr0 RELATED,ESTABLISHED
>  ACCEPT virbr0 virbr0
>  ACCEPT virbr0 any src:192.168.122.0/24
>  REJECT any virbr0
>  REJECT virbr0 any
>
> And after second network is started, we'd have:
>
>  # This one is new:
>  REJECT virbr1 virbr0

Yeah, it's (mostly) this rule that makes the solution complicated.
Essentially you need one of these rules for each pair of virtual
networks, so if you have 10 networks, you'll need 45 rules, for 11
you'll need 55 (in general, for "n" networks, you'll need (n * (n-1) /
2) extra rules).

Aside from the extra packet processing overhead caused by the high
number of rules needed (which may be completely insignificant), and the
necessity to add/remove rules at specific labels rather than just
inserting them all at the beginning, there is the issue that currently
each iptables rule added for a libvirt network is dependent only on that
network, so it's simple to determine which rules to remove when a
network is destroyed. If we start adding rules that are dependent on two
networks, then proper cleanup will be more complicated.

None of that should prevent a properly motivated person from writing the
code to do it, of course!

>  # These ones would be normally at the end, IIRC:
>  ACCEPT any virbr1 RELATED,ESTABLISHED
>  ACCEPT virbr1 virbr1
>  ACCEPT virbr1 any src:192.168.122.0/24
>  REJECT any virbr1
>  REJECT virbr1 any
>  # These are left as they were:
>  ACCEPT any virbr0 RELATED,ESTABLISHED
>  ACCEPT virbr0 virbr0
>  ACCEPT virbr0 any src:192.168.122.0/24
>  REJECT any virbr0
>  REJECT virbr0 any
>
> Cc'ing Laine so that he can weigh in as he has way more knowledge of
> this part of the code =)