This is not a new bug, I'm just dumping my old bug reports
into your new database.

Whoever designed the IP aliases support made the assumption
that they are only used for web servers. In other words, the
initscript add host routes to those aliases.

After several hours of going through all the network
scripts, I used this fix. It's not necessarly great, but
it works. The patch is small enough to be applied by hand,
so I'm not sending a copy by mail.
It checks if there is already a route to the network the
alias belongs to, and adds one if it isn't the case

--- ifup.redhat Sat Aug 29 18:43:24
1998
+++ ifup        Sat Aug 29 19:18:46
1998
@@ -130,7 +130,9
@@

     ifconfig ${DEVICE} ${MACADDR:+hw ether $MACADDR}
${IPADDR} \\
       netmask ${NETMASK} broadcast
${BROADCAST}
-    if [ "$ISALIAS" = no ] ;
then
+
+    # Setup route to aliases if they're on a different
subnet
+    if [ "$ISALIAS" = no -o z"`route -n | sed "s/ .*//" |
grep ${NETWORK}`" = z
 ] ;
then
        route add -net ${NETWORK} netmask ${NETMASK}
${DEVICE}

else
        route add -host ${IPADDR} ${DEVICE}


Note that this will only work if aliased interfaces go
through ifup
This means that I had to patch ifup-aliases to completely
bypass linuxconf because there was no simple way to fix the
problem.

I think that the way you handle aliases is plain broken:
aliases should be normal devices that go through ifup.
However, at alias creation time, the GUI should give the
option of creating a "web style" alias and put a netmask of
255.255.255.255.
Initscripts could then treat the alias like a normal
interface (just like before you modified
/etc/rc.d/init.d/network to remove aliases with a egrep -v
'(ifcfg-lo|:))

Obviously, most of those problems go away with 2.1.x kernels
since the route to the local subnet is created
automatically. I guess that not fixing this and doing a
rewrite for 2.2 would be reasonable.