Arch Linux

I've solved this problem by assume only bond0 exist.
vi /etc/rc.d/network
# add following function
bond_down()
{
for ifline in ${BOND_INTERFACES[@]}; do
if [ "$ifline" = "${ifline#!}" ]; then
eval bondcfg="\$bond_${ifline}"
/sbin/ifenslave -d $ifline $bondcfg || error=1
fi
done
}

# modify the stop) section on line ~ 230-242, change the ifdown section to
R_IFACE=() # use reversed order
for ifline in ${INTERFACES[@]}; do
R_IFACE=("$ifline $R_IFACE")
done
for ifline in ${R_IFACE[@]}; do
if [ "$ifline" = "${ifline#!}" ]; then
if [ "$ifline" = "bond0" ]; then
bond_down
fi
ifdown $ifline || error=1
fi
done

TODO: modify the total network script, all stop should better change the stop order to reversed.

status in 2008-02?

Roman, I know you mentioned doing some bonding work recently. Do you know if this is resolved for you?

Yes, I was doing a backend for web-based configuration system, but that was on CentOS.
Tomorrow on my job I will continue the same work, and I'm going to install Arch on VMware server there to test & try to fix those bonding configurations (VMware Server allows up to 4 ethernet interfaces which is enought to test pretty different bonding configs).

Do we even need the R_IFACE section? Could we not just always call bond_down on shutdown?

I am trying to make sense of the original report, but there doesn't even seem to be one complete sentence in the description. Can anybody explain what this is about?

The way I understand it is that bond devices aren't shutdown properly, making them fail if you restart or whatever. I may be wrong, but I got that by reading the code in the first comment and comparing to the existing code.

Hi, Thomas, the key point is:

It required to shutdown network in REVERSE order when multiple interface attached.(especial the bond)

Is this fixed in the latest initscripts?

First of all, I apologize for the huge delay on this issue,
real life, lack of time and laziness...

Back on topic:
the problem turned out to be very simple, and has nothing to do with real bonding.

suppose we have this in rc.conf (one card - two IP addresses):
eth0="eth0 ..."
eth1="eth0:1 ..."
INTERFACES=(eth0 eth1)

`/etc/rc.d/network stop` will try to stop eth0 *_and_eth1_* which does not exist,
(if you run ifconfig with the configuration above - you'll see eth and eth:0 but no eth1)
thus /etc/rc.d/network prints FAIL, though eth0 and eth0:1 are stopped correctly.

The solution is either don't FAIL on inability to remove existing device,
or improve the way this type of configurations is defined:
eth0=("eth0 ..."
"eth0:1 ...")

There are also some issues with real bonding (I guess Daniel experienced two different issues at the same time),
I think slaves should be unbonded first before shutting down the master interface.

Yes, Roman, I mean, we should release the interfaces with INVERSE order.
Otherwise in my bonding situation. It can only 'network start' but 'network stop'

My rc.conf:

eth1="dhcp"
eth1_1="eth1:1 192.168.1.200 netmask 255.255.255.0 broadcast 192.168.1.255"
eth1_2="eth1:2 192.168.1.201 netmask 255.255.255.0 broadcast 192.168.1.255"
eth1_3="eth1:3 192.168.1.202 netmask 255.255.255.0 broadcast 192.168.1.255"
INTERFACES=(eth1 eth1_1 eth1_2 eth1_3)

The problem still exists in initscripts whether it be bonding or ip aliasing. Has there been any progress on this at all?

Thanks,

Can someone have a look at this patch:

--- network-orig 2009-12-06 13:35:15.000000000 -0500
+++ network 2009-12-06 20:19:41.000000000 -0500
@@ -19,8 +19,6 @@
return 1
fi

- /sbin/ifconfig $1 up
-
wi_up $1 || return 1

eval ifcfg="\$${1}"
@@ -65,7 +63,11 @@
fi
fi
# Always bring the interface itself down
- /sbin/ifconfig ${1} down >/dev/null 2>&1
+ # Ignore aliases
+ `echo ${1} | grep -q '_'`
+ if [ $? -gt 0 ]; then
+ /sbin/ifconfig ${1} down >/dev/null 2>&1
+ fi
return $?
}

I'd like to get an answer on why we run /sbin/ifconfig $1 up without doing any tests at all in the ifup function?

Also I think its a good idea to ignore aliased interfaces all together since bringing down the parent interface will automatically remove the aliased interfaces.

Thanks

We are considering limiting the scope of the network initscript to not include this usecase. Is there any reason you cannot use netcfg instead?