AIR Wiki : LogicalInterfaceNames

HomePage :: Categories :: PageIndex :: RecentChanges :: RecentlyCommented :: Login/Register

How to reorder or rename logical interface names in Linux


One of the problems of Linux is that the order of the network interfaces is unpredictable. Between reboots it usually stays the same, but often after an upgrade to a new kernel or the addition or replacement of a network card (NIC) the order of all network interfaces changes. For example, what used to be eth0 now becomes eth1 or eth2 or visa versa.

Obviously there is some logic to which network interface gets which name, but Linux documentation states that this may change and no user or program should ever assume anything about this. This is annoying, in particular if your management interface is at eth1 at one node in a cluster and at eth2 in another node of the same cluster (which we have experienced). I personally like to have my (primary) management interface always to be eth0.

Thankfully, there are ways to achieve this. They can be divided in four methods:
1. Order the network interfaces based on physical properties of the NIC. (e.g. the physical location in the machine)
2. Order the network interfaces based on the MAC address of the NIC.
3. Order the network interfaces based on the driver of the NIC.
4. Order the network interfaces based on the physical location of the NIC in the computer

So you have to pick a method that suits you. I recommend either to use ifrename (based on physcial properties, especially useful if you often change network cards in your hosts) or writing a udev rule (based on the MAC address). However, I listed the other methods as well. Be aware that the last two methods mentioned in this article are only for the masochistic (you will scream and shoot to get those to work).

Note: Linux kernels up to 2.4 did only probe for the first Ethernet card, ignoring other NICs. We assume you use a 2.6 or higher kernel or already fixed this, for example by specifying ether=0,0,eth1 as kernel parameter.

Based on the physical properties


Perhaps the most elegant way to name the ethernet NIC is to do so based on their physical properties, like link speed and port type.

Using the ifrename tool

Ifrename is a tool specifically designed to (re)name network interfaces based on characteristics like MAC address (wildcards supported), bus information, and hardware settings. It uses a control file (/etc/iftab) to specify rules about how the interfaces will be named. (thanks to Matt Baron for this tip.)

# Example /etc/iftab file
eth2           mac 08:00:09:DE:82:0E
eth3           driver wavelan interrupt 15 baseaddress 0x390
eth4           driver pcnet32 businfo 0000:02:05.0
# wildcard name: pick the lowest available name of air0, air1, air2, etc.
air*           mac 00:07:0E:* arp 1


Using the ethtool and ip programs

It is possible to check the NIC properties using the ethtool program, and to change the name using the ip program (thanks to Jollynn Schmidt for this tip):

if ethtool eth0 | grep -q "Port: FIBRE"; then
	   ip link set dev eth0 name not_eth0
	   ip link set dev eth1 name eth0
	   ip link set dev not_eth0 name eth1
fi


The disadvantage of ethtool is that it can only be run by root, even when you're only using it to query for information. Though this is a minor annoyance of ethtool, it doesn't matter in this case, since you want to set a device name and thus need to be root anyway.

Based on the MAC address


Secondly, it is also possible to name the network interface based on the MAC address of each NIC. The advantage is that it is possible to use this method if you have two NICs which use the same driver (unlike the next method: based on driver).

First, you must determine the MAC address of your interfaces. You can do this locally on a machine running
ifconfig -a

The MAC address is listed as "hwaddr" (hardware address). Alternatively, you can even determine MAC addresses remotely using ping and /sbin/arp.

There are three ways to map the MAC address to the logical interface name. Either by using the udev rules, with the get-mac-address.sh script, or by using the nameif program.

The udev method should work on all recent Linux distributions, and is recommended. The get-mac-address.sh script and the nameif program are know to work with Debian, while on Red Hat, you can change the interface configuration file.

Using udev rules
udev replaced devfs in Linux 2.6. First make sure that your Linux system has udev installed, rather then devfs. If you have a /etc/udev directory, but not /etc/devfs directory, you are probably fine. If not, be aware that changing your kernel from devfs to udev is possible, but is not just a matter of adding a new module. Perhaps for now, another method is easier for you.

Now that you have udev, it is rather simple. You only need to create a udev rule mapping the MAC address to the interface name. Store this in a file inside the /etc/udev/rules.d/ directory:
KERNEL=="eth?", SYSFS{address}=="00:37:e9:17:64:af", NAME="eth0"  # MAC of first NIC in lowercase
KERNEL=="eth?", SYSFS{address}=="00:21:e9:17:64:b5", NAME="eth1"  # MAC of second NIC in lowercase


Most distibutions already come with an example config file for you. E.g. /etc/udev/rules.d/network-devices.rules or /etc/udev/rules.d/010_netinterfaces.rules. More information can be found at http://www.reactivated.net/writing_udev_rules.html or http://www.debianhelp.co.uk/udev.htm. (Thanks to Casey Scott and Ron Hermsen for the pointers.)

Another rule I've seen is:
SUBSYSTEM=="net", DRIVERS=="?*", ATTRS{address}=="00:16:3e:00:02:00", NAME="eth0"
SUBSYSTEM=="net", DRIVERS=="?*", ATTRS{address}=="00:16:3e:00:02:01", NAME="eth1"


I'm not sure about the difference between these rules. Information is welcome (please leave a comment below)


Using the interface configuration file
If you run a Red-Hat-based distribution, you can simply add the MAC addres in the interface configuration file /etc/sysconfig/network-scripts/ifcfg-eth0:

DEVICE=eth0
HWADDR=00:37:e9:17:64:af


You can give it any DEVICE name you want, like DEVICE=ethmgmt, as long as you remember to rename the config file:
/etc/sysconfig/network-scripts/ifcfg-ethmgmt

Source: http://forums.serverwatch.com/showthread.php?t=18476

Using the get-mac-address.sh script
Another solution is to use the get-mac-address.sh script to map interface names by MAC address. On Debian, this script is distributed as part of the ifupdown package (in /usr/share/doc/ifupdown/examples/get-mac-address.sh). Copy this script to a saner place (e.g. /usr/local/bin), and you can setup /etc/network/interfaces in this manner:

auto lo eth0 eth1

iface lo inet loopback

mapping eth0 eth1
	    script /usr/local/bin/get-mac-address.sh
	    map 00:37:E9:17:64:AF netA
	    map 00:21:E9:17:64:B5 netB

iface netA inet static
	    address etc...

iface netB inet static
	    address etc...


Source: https://www.gelato.unsw.edu.au/archives/gelato-technical/2004-February/000334.html

The disadvantage of this method is that defines a mapping, rather then changing the actual logical interface name.

Using the nameif program

Alternative to the get-mac-address.sh script, you can also use the slightly more convenient nameif program, which is distributed as part of the net-tools package on Debian.

The advantage of nameif is that you can specify the interface names in the /etc/mactab file:
ethmgnt 00:37:E9:17:64:AF
ethwireless 00:21:E9:17:64:B5


It is not possible to rename an interface to a name of an existing interface. So you can't rename eth1 to eth0 as long as eth0 still exists. It is possible to still swap the names eth0 and eth1 by using a temporary name (e.g. first rename eth1 to ethfoo, then eth0 to eth1 and finally ethfoo to eth0). Note that this method may lead to problems if you use common names such as eth0 and eth1. If you upgrade a kernel, the names may be different than you expected, and you may rename a NIC to eth0 while eth0 still exists, leading to name collisions. Therefor, it is recommended to use other names like "ethmgmnt", "ethwired", "ethwireless" and "eth10ge", as shown in the example above.

Based on the driver


Warning: This only works if the driver is available as a loadable module. Not if it is built into the kernel.

This is a relative easy method, since it does not rely on external scripts. The idea is to just load the kernel module for your eth0 interface before the modules for other network cards.

First of all, you must determine which driver is used for each network card. Thankfully Linux does have a system to load the appropriate driver automatically, based on the PCI ID of the network card. Unfortunately, there is no single command to simply get the driver (and other information like the link speed) based on just the interface name in Linux. Your best bet is to look for kernel messages:
dmesg | grep eth

This should give you a good estimate of the driver name. You can verify if the name indeed does exist and is loaded:
lsmod


Note:
lsmod gave:
e1000           84868  0
tg3             70816  0

However, the 0 indicates that these drivers are not controlling any device! That is strange, since modprobe -r tg3 and modprobe -r e1000 do disable the network cards. Apparently, this is a flaw in lsmod.

Note that running modprobe tg3 en then modprobe e1000 does bring them up in the correct order, with the correct interface names. This is a good check if this approach (using the driver to decide the interface name) can work.

Red Hat
In Red Hat, if the driver is called "tg3" (the Tigon driver), you simply specify the network name by adding this enty in /etc/modules.conf:
~alias eth0 tg3


Debian
On a Debian system, /etc/modules.conf is generated automatically and should not be edited directly. Instead, create a file in the subdirectory /etc/modules/ (do not use /etc/modprobe.d/, that seems out-of-date). For example, create the file /etc/modutils/interfaces and add the appropriate modules. For example:
alias eth0 tg3
alias eth1 e1000

Next, update /etc/modules.conf by running:
update-modules


Alternative method: I have encountered scenario's where the kernel did already load the modules for the drivers, even before /etc/modules.conf was read. The result was that in effect, the specification in /etc/modules.conf was ignored, and this method did not work. As an alternative, it is possible to also list the drivers, in the appropriate order, in /etc/modules (thus not /etc/modules.conf):
tg3
e1000

The result will be that the tg3 driver is loaded before the e1000 kernel. Since /etc/modules only exists for Debian, this trick will most likely not work for other distributions.

Based on the physical location in the computer


Warning: This only works if the driver is built into the kernel, not as a loadable module.

Note: It is relatively hard to get this to work, and we encountered problems with it. The other methods are recommended.

It is possible to name the network interfaces based on the interrupt (IRQ) and memory address. This should work if you have network cards in PCI busses, and it involves appending the proper parameters to the "ether=" or "netdev=" kernel parameters.

First of all, you can detect the PCI slot of the devices using
lspci -v
This is reported to fail sometimes for certain cards. Now, write down the IRQ and IO (memory) address of each network card, and use this information to specify the interface name in your LILO or GRUB configuration file.

For LILO, you need to add an add line to the appropriate boot configuration. For example:
append="netdev=irq=21,io=0x2040,name=eth0 netdev=irq=20,io=0x3000,name=eth1 netdev=irq18,io=0x2000,name=eth2"

Under grub, it can just be listed as parameter. e.g.:
kernel /boot/vmlinuz netdev=irq=24,name=eth0


More Information
http://www.tldp.org/HOWTO/Ethernet-HOWTO-8.html
http://www.tldp.org/HOWTO/BootPrompt-HOWTO-11.html


Written by Freek Dijkstra. Licensed under public domain. (That is, feel free to modify, redistribute, cripple, or even sell it as your own work, and there is no need to mention the source, even though you are of course welcome to do so.)

Categories
CategorySysAdmin
CategoryProcedures
 Comments [Hide comments/form]
I had to change the name of an interface that was used to netboot.

Linux doesn't allow changing the name of an active interface and bringing down the netbooted interface meant losing the ability to access the very programs needed to perform the name change. My solution (excerpted below) included a chroot. The code below uses "nameif" but could almost as easily used "ifrename" and probably even "udev". This code runs on Fedora 7.

#!/bin/sh
#
# Use "chroot" to allow changing the name of an interface that
# has been used to net boot to (such that bringing down that
# interface will cause loss of ability to access any of the
# operating system programs)
#
dir=/var/tmp/nameifRoot # location on local machine
mac=00:01:02:03:04:05 # change to mac address of interface to change
oldname=eth0 # change to current name of interface
newname=control # change to what you want the interface called
#
# Figure out what is needed in the chroot area
#
needed=""
#
# Use "ldd" to figure out what libraries are needed for the given
# executables. The executables are returned with a trailing ':',
# hence the sed command to strip all the colons (this colon
# stripping is simple and expediant and will fail miserable if a
# colon shows up in the path of one of the needed files).
#
lddout=`ldd /sbin/nameif /sbin/ifconfig /bin/sh | sed 's/://'`
for w in ${lddout}
do
#
# ldd spits out lots of words that are not paths,
# we decide that the only thing that are paths are
# words that start with a '/'
#
c1=`echo $w | cut -c1`
if [ "$c1" = "/" ]
then
needed="${needed} $w"
fi
done
#
# Know now what is needed, do the copying
#
for l in ${needed}
do
mkdir -p ${dir}/`dirname ${l}`
cp $l ${dir}/$l
done
#
# Make the "mactab" file for "nameif" to use.
#
mkdir -p ${dir}/etc
echo "${newname} ${mac}" > ${dir}/etc/mactab
#
# Mount "proc" (nameif and ifconfig use it)
#
mkdir -p ${dir}/proc
mount -t proc proc ${dir}/proc
#
# Make a /dev/null (mostly so we don't have to see the errors)
#
mkdir -p ${dir}/dev
mknod -m 666 ${dir}/dev/null c 1 3
#
# Make the program to run. Bring down the ethernet interface,
# do the nameif, and then bring everything back up again.
#
cat <<@eof > ${dir}/go
ifconfig ${oldname} down >& /dev/null
nameif
ifconfig ${newname} up >& /dev/null
ifconfig ${oldname} up >& /dev/null # just in case nameif failed
@eof
chmod 700 ${dir}/go
#
# Do the renames
#
(chroot ${dir} /go) &
#
# what happens when the network goes away? Apparently, wait will
# wait for it to come back :-)
#
wait
#
# Cleanup
#
umount ${dir}/proc
rm -rf ${dir}
echo "done"
-- MichaelHeyman (2007-10-29 14:49:57)
In the section "Using udev rules" I can tell you what difference they make. I have a CentOS kickstart setup that I use for Xen Dom0's. In this setup I have bonded interfaces as well as VLAN trunks. It took me a few days to figure out why I could never get the VLAN device names to correctly initialize.

Once I setup UDEV debug logging I saw that udev was trying to rename the VLAN devices with the udev rule I setup in kickstart for persistent MAC to dev naming. Since the VLAN dev does inherit the MAC address from the parent, in my case a bond device which is in turn getting MAC from the underlying eth device, and they fall in the "SUBSYSTEM=="net"" catagory it was matching my rules.

I added "KERNEL=="eth?*"" as the first parameter and it solved my issue. The only case I can see this not working is if you put VLAN interfaces directly on eth devices. In that case adding "SYSFS{features}=="0x0"" with the hex set to match anything BUT 0x0 might work. It looks to me like {features} is always 0x0 on VLAN devices. All of my bond devices and real devices have something else in that value.

This is my working rules file:

## Created by Kickstart to keep network interfaces in an expected order

## Xen Bridge aggr eth0 interface
## eth0: Tigon3 [partno(BCM95704A7) rev 2003 PHY(5704)] (PCIX:100MHz:64-bit) 10/100/1000Base-T Ethernet 00:e0:81:2e:d1:80
KERNEL=="eth?*", SUBSYSTEM=="net", SYSFS{address}=="00:e0:81:2e:d1:80", NAME="eth0"

## Xen Bridge aggr eth1 interface
## eth1: Tigon3 [partno(BCM95704A7) rev 2003 PHY(5704)] (PCIX:100MHz:64-bit) 10/100/1000Base-T Ethernet 00:e0:81:2e:d1:81
KERNEL=="eth?*", SUBSYSTEM=="net", SYSFS{address}=="00:e0:81:2e:d1:81", NAME="eth1"

## Xen Bridge aggr eth2 interface
## e1000: eth2: e1000_probe: Intel(R) PRO/1000 Network Connection
KERNEL=="eth?*", SUBSYSTEM=="net", SYSFS{address}=="00:04:23:08:68:fc", NAME="eth2"

## Xen Bridge aggr eth3 interface
## e1000: eth3: e1000_probe: Intel(R) PRO/1000 Network Connection
KERNEL=="eth?*", SUBSYSTEM=="net", SYSFS{address}=="00:04:23:08:68:fd", NAME="eth3"

## Xen Bridge aggr eth4 interface
KERNEL=="eth?*", SUBSYSTEM=="net", SYSFS{address}=="00:e0:81:2e:d1:46", NAME="eth4"


If anyone is interested, this is my current kickstart script that generates all of this:

http://www.themelon.org/ks/equilibrium.themelon.org.cfg
-- TheMelon (2009-03-10 01:42:04)
There's two differences between the rules "KERNEL=="eth?", SYSFS{address}==..." and "SUBSYSTEM=="net", DRIVERS=="?*", ATTRS{address}==..."
There's the "SYSFS{address}" and "ATTRS{address}" which have *exactly* the same effect - to identify the hardware by it's address. Both are valid ways of getting the information, but in Debian at least udev complains that "SYSFS{}= will be removed in a future udev version, please use ATTR{}= to match the event device", which is fairly self explanatory.

There's also the 'KERNEL=="eth?"' and 'SUBSYSTEM=="net", DRIVERS=="?*"'. In the first rule, 'KERNEL=="eth?"' it will match any hardware whose name is eth0 to eth9, and the MAC address matches. This presents a problem if the hardware is a USB wireless card, which is often named "wlan0", as it wont be matched with that rule. The second rule however matches any hardware which is a network interface - the "SUBSYSTEM" key "net" matches any network interface*. So, the second rule will match a network card if it is called "eth0", "eth1", "wlan0", or whatever, and then match the specific MAC address of the card. The 'Drivers=="?*"' is (I believe, but I might be wrong) unnecessary, so I don't have it (and it works for me without it).

My current udev rule for a USB stick, which I use with a USB-wireless adaptor (it's easier to work with known hardware) so I can always get to it at "usb-wifi" is:

SUBSYSTEM=="net", ATTRS{address}=="08:10:74:53:35:4d", NAME:="usb-wifi"

This way no matter what other network hardware is in the machine (be it another USB wireless, or onboard ethernet, or...) my USB wireless stick will always be matched - it's always in the network subsystem, and its MAC address is always the same - and named "usb-wifi". The ":=" means "don't let this be changed by another rule" so if (for some strange reason) another rule running later wants to change the name of this network card, it can't.

* I suggest you go to a console on your machine and type "ls /sys/class/net" it will list all the network interfaces. The "/sys" directory contains all the information you can use to match hardware, do a "ls /sys/class/net/eth0" (or replace eth0 with your network interface) and you'll see several files you can "cat" to get information on, such as "address" for the MAC address.
-- PenguinMe (2010-01-15 22:45:11)
Valid XHTML 1.0 Transitional :: Valid CSS :: Powered by Wikka Wakka Wiki 1.1.6.0
Page was generated in 0.0474 seconds