linux - When IP aliasing how does the OS determine which IP address will be used as source for outbound TCP/IP connections?

I have a server running Ubuntu Server with four IP addresses aliased on a single NIC.

eth0       192.168.1.100
eth0:0     192.168.1.101
eth0:1     192.168.1.102
eth0:2     192.168.1.103

(Using 192.168.x.x for sake of example, assume these are NAT-ed to a range of public IP addresses)

One of our clients publishes their inventory via FTP, so we log in nightly to download a large file from their server. Their firewall expects our (passive) FTP connection to be made from 192.168.1.100.

Given that my server logically has four IP addresses on a single adapter, how does the operating system determine which IP address is used as source for outbound TCP/IP connections?

Let's say I ssh into my server on 192.168.1.101 and run FTP interactively. Will the outbound TCP/IP connection use 192.168.1.101 because the OS knows that's the interface over which my shell is connected?

What if the FTP task is run non-interactively via a cron job where there is no shell?

As you can probably tell, this has me quite confused, so I hope my questions have at least
made sense.

Edit

To clarify why I'm asking -- I haven't made any changes to the routing table and it actually lists 'eth0' as the IFace for the 0.0.0.0 routes. However, all indications are that it is actually using eth0:0 as the source.

Destination    Gateway         Genmask         Flags Metric Ref    Use Iface
192.168.1.0    0.0.0.0         255.255.255.0   U     0      0        0 eth0
0.0.0.0        192.168.1.1     0.0.0.0         UG    100    0        0 eth0

I can fiddle with the routing table or have our client change their firewall rules to get the behavior I need, but I'm trying to gain a little insight into how this works to know if there's a bug in the OS or just my naive understanding of how all the pieces fit together.

Thanks

Answer

By default, on Linux, if an interface has multiple addresses that are on different subnets, traffic destined for the respective subnets will have the proper source IP. That is, if eth0 has two addresses 192.168.1.1/24 and 10.1.1.1/8, then traffic to anything on the 10.0.0.0 subnet will have source 10.1.1.1, and traffic to anything on the 192.168.1.0 subnet will have source 192.168.1.1. You can also assign source addresses explicitly in this case by using the "src 1.2.3.4" option to "ip route".

In your case, though, all your addresses are on the same subnet, so the "primary" one (as revealed by "ip addr list dev eth0") is used as the source IP for traffic exiting on that interface. I think it's possible to control the source IPs in this case just using "ip route", but I've found it easier to use iptables to rewrite the source addresses for traffic of interest.

If you want to force a specific source address to be used for specific destinations, you can do it with a SNAT rule:

iptables -t nat -I POSTROUTING -o eth0 -d dest-IP-or-net/mask -s primary-IP-of-eth0 -j SNAT --to-source desired-source-IP

So if your "primary" eth0 IP is 192.168.100.1, but you want traffic to 1.2.3.4 to have a source of 192.168.100.2, then do this:

iptables -t nat -I POSTROUTING -o eth0 -d 1.2.3.4/0 -s 192.168.100.1 -j SNAT --to-source 192.168.100.2

Note that the "-s 192.168.100.1" is important: it prevents forwarded traffic's source addresses being rewritten by this rule.

If you are going to implement complex network configurations on Linux, you should read the Linux Advanced Routing and Traffic Control documentation, http://lartc.org