Stale route not flushed/updated

[Arnaud Lacombe <lacombar%gmail.com@localhost>]

Hi all,

I got a routing problem this afternoon with a soekris box running 5.0RC3.

A short description of the problem is that the kernel was re-using an
old local source IP address in an ICMP packet. Flushing routes and
recreating them solved the problem.

A longer description of the problem is the following: an interface on
the soekris has 4 IP address:

$ cat /etc/ifconfig.vr0
inet 172.16.0.1 netmask 0xffffff00
alias 172.16.1.1 netmask 0xffffff00
alias 172.16.2.1 netmask 0xffffff00
alias 172.16.3.1 netmask 0xffffff00

and 4 different routes associated:
route add 10.1.0.0/24 172.16.0.200
route add 10.1.1.0/24 172.16.0.200
route add 10.1.2.0/24 172.16.0.200
route add 10.1.3.0/24 172.16.0.200

as expected, the route to 10.1.3.1 picks 172.16.0.1 as the local source address:
# route get 10.1.3.1
   route to: 10.1.3.1
   destination: 10.1.3.0
   mask: 255.255.255.0
   gateway: 172.16.0.200
   local addr: 172.16.0.1
   interface: flags: <UP,GATEWAY,DONE,STATIC>

Now, if I clean-up the address on vr0:
# ifconfig vr0 delete 172.16.0.1
# ifconfig vr0 delete 172.16.1.1
# ifconfig vr0 delete 172.16.2.1

# ifconfig vr0
[...]
inet 172.16.3.1 netmask 0xffffff00 broadcast 172.16.3.255

I still have the old route entry pointing to 10.1.3.1:
# route get 10.1.3.1
   route to: 10.1.3.1
   destination: 10.1.3.0
   mask: 255.255.255.0
   gateway: 172.16.0.200
   local addr: 172.16.0.1
   interface: vr0 flags: <UP,GATEWAY,DONE,STATIC>

At this point, trying to ping 10.1.3.1 fails as expected:
# ping 10.1.3.1
PING 10.1.3.1 (10.1.3.1): 56 data bytes
ping: sendto: No route to host

However, if I change the netmask associated with 172.16.3.1 to a /22,
allowing the access to 172.16.0.200. Trying to ping 10.1.3.1 generates
an ICMP packet with the wrong source IP:

09:05:17.238201 IP 172.16.0.1 > 10.1.3.1: ICMP echo request, id 31459,
seq 5, length 64

The only point to recover from this situation is to flush the route to
10.1.3.0/24 and re-create it.

From my understanding, the route entry with 172.16.0.1 as a local
source to 10.1.3.0/24 should be flushed as soon as the interface no
longer have the IP address associated.

comments ?

 - Arnaud

ps: btw, is there a way to tell the routing code to use a specific
source address depending of the destination network. In the above
setup, I'd like to have the 4 /22 (/24 currently, I know) using
172.16.0.200 as a gateway but picking different local source address
depending of the destination.