Re: How to override the default source ipv4 address on packets originating from a router

[Brian Burch <brian@pingtoo.com>]

On 21/12/13 12:53, Brian Burch wrote:
> On 20/12/13 18:46, Brian Burch wrote:
>> On 20/12/13 13:05, Joel Gerber wrote:
>>> Your problem is one I've come up against a number of times in the
>>> past. You are definitely on the right route using "policy routing"
>>> (otherwise known as utilizing IP Rules), but the snag you're hitting
>>> is that the source address of a locally generated packet isn't chosen
>>> until a destination route has already been selected. Using the from
>>> statement in an ip rules table is only useful when you're routing
>>> packets from other subnets.
>>
>> What a quick and helpful reply! Not only did my problem motivate you to
>> respond, but I am very relieved to find the answer isn't just RTFM.
>>
>> Your explanation about the timing for kernel selection of the source
>> address is crucial and now my failed experiments all make sense. Your
>> suggested solution also makes sense, although it is counter-intuitive...
>> at least to me!
>>
>>> Now, just so you know, the local table should normally be left alone.
>>> It's used for locally connected subnets only. It is given the highest
>>> priority on purpose... don't mess with it! :)
>>
>> I read somewhere (possibly the howto guide or the ip-cref) a statement
>> that the kernel treats this table as read-only. My system is running
>> ubuntu 13.04 with a custom non-pae 3.8.0 kernel, so it is reasonably
>> modern. When I thought more about the purpose of this table, I decided
>> there would be no point in trying to modify it, even if that was
>> possible.
>>
>>> There are a few methods you could try here, but I'm only going to
>>> recommend one of them. Have your main routing table hold the routes
>>> that your local machine will use, and then have a separate table for
>>> all routed hosts. Then, use ip rule statements to force any traffic
>>> coming in from a routed host into that routing table.
>>>
>>> IE: on firewall-router (assuming eth1 is facing your static subnet
>>> hosts):
>>>
>>> # ip rule list
>>> 0:  from all lookup local
>>> 32765:  iif eth1 lookup static-hosts
>>> 32766:  from all lookup main
>>> 32767:  from all lookup default
>>>
>>> Have table main include all of the routes your local firewall-router
>>> should use, and have static-hosts contain all of the routes that your
>>> static-subnet-hosts should use.
>>
>> Great! I understand the principle you propose and will try to implement
>> it over the weekend. I will update this thread when I have something to
>> report.
>
> Excellent news!! My router/firewall is currently collecting 200+ updated
> packages from the repository server!
>
> The solution was achieved using ONLY policy routing, based on Joel's
> suggested approach.
>
> However, I hit some fairly nasty snags along the way and had to do a lot
> of trial and error fiddling to make it work. I am certain my working
> solution can be made more elegant, but that will take a couple of hours
> pen-and-pencil time to map out the current routing table entries in
> selection order and then perform some experiments.
>
> I don't have time to do it at the moment, but will get started over the
> next couple of days. If I get stuck, I will post the "current best"
> configuration and ask for help to improve it.

<snip>

I tried to improve my "current best" definitions, but my efforts to 
remove apparent redundancy have all failed. Here is the setup:

brian@chenin:~$ ip address show
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN
     link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
     inet 127.0.0.1/8 scope host lo
     inet6 ::1/128 scope host
        valid_lft forever preferred_lft forever
2: eth1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast 
state UNKNOWN qlen 1000
     link/ether 00:15:8a:01:91:e3 brd ff:ff:ff:ff:ff:ff
     inet 217.154.193.209/28 brd 217.154.193.223 scope global eth1
     inet6 fe80::215:8aff:fe01:91e3/64 scope link
        valid_lft forever preferred_lft forever
3: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast 
state UNKNOWN qlen 1000
     link/ether 00:40:63:f5:a5:72 brd ff:ff:ff:ff:ff:ff
     inet 172.16.101.2/24 brd 172.16.101.255 scope global eth0
     inet6 fe80::240:63ff:fef5:a572/64 scope link
        valid_lft forever preferred_lft forever


brian@chenin:~$ ip rule list
0:	from all lookup local
32764:	from 217.154.193.209 lookup CHENIN_TO_INTERNET
32765:	from all iif eth1 lookup SUBNET_217
32766:	from all lookup main
32767:	from all lookup default


brian@chenin:~$ ip route list table CHENIN_TO_INTERNET
163.1.221.67 via 172.16.101.1 dev eth0


brian@chenin:~$ ip route list table SUBNET_217
default via 172.16.101.1 dev eth0


brian@chenin:~$ ip route list table main
163.1.221.67 dev eth0  scope link  src 217.154.193.209
169.254.0.0/16 dev eth0  scope link  metric 1000
172.16.101.0/24 dev eth0  proto kernel  scope link  src 172.16.101.2
217.154.193.208/28 dev eth1  proto kernel  scope link  src 217.154.193.209


I know Joel's suggestion did not include my CHENIN_TO_INTERNET rule, but 
without it chenin simply times out on its http gets.

I also tried keeping CHENIN_TO_INTERNET, and removing the explicit route 
in the main table, but that fails too, although I wasn't very surprised.

I have tried trimming these rules and routes to discover the minimum 
requirement, but this seems to be the smallest working subset.


Here is my theory to account for what is happening:

1. The SUBNET_217 default route MUST be able to handle ANY traffic 
arriving on eth1. When I changed the rule to match on just 
217.154.193.208/28, the forwarded traffic became very glitchy, with poor 
response times and some dropped tcp sessions.

2. The route to 163.1.221.67 in the main table is required to assign the 
eth1 source address 217.154.193.209 to its outbound packets. This works 
because the kernel cannot find any other matching routes from chenin to 
163.1.221.67 (which might encourage it to assign eth0's 172.16.101.2 as 
the source address).

4. Having assigned the 217.154.193.209 source address, the packet must 
be routed to the next hop. The CHENIN_TO_INTERNET rule matches the 
explicit source host address at a higher priority than the SUBNET_217 
rule, so 172.16.101 is selected.

5. The main table has a route to 172.16.101.1 via eth0, so the packet is 
sent on its way.


That seems to be three passes through the policy routing tables. Am I 
correct?

I still feel it should be possible to improve the solution, but I don't 
understand what is going on well enough to see how. I would be very 
grateful if anyone can enlighten me.

Thanks,

Brian