Sane network interface management with Hotplug ---------------------------------------------- INTRODUCTION ------------ In the old days all wireless cards were managed by the excellent Pcmcia subsystem and its rich configuration scripts, and life was good. Then came the wireless PCI cards, then the wireless USB dongles. Some unification was needed, and rather than adapt the Pcmcia subsystem for PCI and USB, it was decided to create the much simpler Hotplug system. The USB subsystem already uses Hotplug. The Pcmcia subsystem is migrating to it : CardBus cards (32 bits) already use Hotplug, whereas Pcmcia cards (16 bits) still use the old Pcmcia scripts. The Hotplug system is still in its infancy, but already shows some good promise. Most users are disappointed at first by its apparent lack of features compared to the Pcmcia scripts. In this document, we will show how to fully exploit the Hotplug system and try to implement the equivalent of all the functionality of the Pcmcia scripts. ASSUMPTIONS ----------- The target audience of this document is mostly power users and distribution maintainers, but it should give enough clues to help newbies. You should have read and understood DISTRIBUTIONS.txt. The procedures described here are more advanced than the simple configuration described in DISTRIBUTIONS.txt. The main focus is of course on removable wireless interfaces, but we will to talk about network interface management in general, so this should apply also to built-in Ethernet cards. PROBLEM STATEMENT ----------------- Let's assume a Linux system and two or more network devices, Device A and Device B. Those devices may be built-in or removable, they may be present or absent from the system at any time, and they may activated in any particular order. The user wants to assign Configuration A to Device A and Configuration B to Device B, without the possibility that Device A gets assigned Configuration B. Different users may have different definitions of what is Device A. For some, it's a specific instance of a specific hardware, for others any hardware that meets some criteria (a wireless card, an Ethernet card). The user may also want to have multiple configurations for a given device such that the chosen configuration depends on various factors, just as with the old Pcmcia schemes. Device A may need Configuration A1 or Configuration A2 depending on those factors. By default, all network interfaces are created using default interface names (starting at "eth0" and going up). I call that the "all my cards are eth0" problem : im most distributions, "eth0" points to a single fixed configuration in the configuration database. Clearly, this won't satisfy our requirements. EXAMPLE SYSTEM -------------- The distribution I use is Debian 3.0, and some parts of what I say here will be specific to it. However, it should be easy to translate this material to other distributions and I welcome additions to this document. The example system is as follows : o Linux 2.6.X SMP kernel with hotplug support o Fully modular system (all network drivers as modules) o PCI Ethernet card : AMD PCnet LANCE (pcnet32 - eth4) o PCI Ethernet card : HP 100VG J2585B (hp100 - eth2) o ISA Wireless card : Old AT&T Wavelan (wavelan - eth3) o ISA-Pcmcia bridge : VADEM VG-469 (i82365 - slot 0) o PCI-CardBus bridge : Ricoh RL5c475 (yenta_socket - slot 2) o Pcmcia 802.11 card : Aironet 350 (airo_cs - eth0) o Pcmcia 802.11 card : Lucent Orinoco (orinoco_cs - eth0) o CardBus 802.11 card : SMC 2835W (prism54 - prism0) This system just happens to be my Linux development box. It has enough interfaces to make it interesting. All the examples I present in this document are extracted from this system. BASIC CONCEPTS -------------- Most of the concept and tricks presented here are not really new. The main contribution is to integrate them. 1) Removable network interfaces are managed by Hotplug (Pcmcia, CardBus, USB...). We can't assume that those interfaces are always present in this system and available at boot time (Pcmcia cards were not made to be soldered in the Pcmcia slot). Therefore Hotplug is the way to go. 2) Built-in PCI and ISA cards are managed by the init scripts, as they have always been. The ISA subsystem will never have Hotplug support, and hotplug is not necessary for PCI cards. 3) Built-in devices that are disable most of the time should be enabled manually by the user. Therefore both Hotplug and the init scripts should ignore those devices by default. 4) (1), (2) and (3) must be compatible on the same system and play nice with each other. 5) A well defined and consistent network interface name is assigned to each network hardware interface using 'ifrename'. Device A is always named 'ethA' (or whatever name you like such as 'mynetworkcard'). 6) No interface is called 'eth0' (or 'wlan0'). Any unknown device would be 'eth0', so known devices should be called something else. 7) Multiple configurations for a single interface (schemes) are managed by the ifup/ifdown subsystem. CONFIGURATION FROM INIT SCRIPTS ------------------------------- It may seem paradoxical, but before setting up Hotplug, we need to make sure that the initialisation of network cards via init scripts is done properly and doesn't get in the way of the Hotplug subsystem. The configuration of network cards via init scripts is the traditional way networking is initialised in Linux. The advantage of this method is that it's very well documented and understood, and has not changed much over the years. Unfortunately, it doesn't adequately support removable cards. The init scripts perform the following 3 functions in order : 1) Load necessary driver modules 2) Rename interface to name chosen by the user 3) Configure those network interfaces 1) Applicability ---------------- Configuration from init scripts is applicable to any built-in network interface (ISA, PCI...), i.e., interfaces available at boot time and that will never be removed from the system. The Hotplug subsystem also has the ability to configure some of the built-in network interfaces, such as PCI cards. However, there is a class of devices that will never have Hotplug support, such as ISA and EISA cards. 2) Loading driver modules (if/as needed) ---------------------------------------- Most distributions build the kernel drivers as modules. This modular setup allows to minimise the amount of memory used by the system and the flexible loading/unloading of drivers. You can also compile your kernel with static drivers (non-modular). In that case, the driver will always be available in the kernel, you don't need to configure the module subsystem, so you can skip directly to the next section. There are 3 alternatives to manage device drivers as modules. 1) Some distributions have an explicit list of modules that are loaded at boot time. If you want to use that feature you need to check the documentation of your distribution. 2) Some system, such as Hotplug, Discover or Kudzu, can scan the various buses of the PC and load the appropriate drivers. This is mostly configuration-free, but may not support all devices and may load unnecessary modules. 3) The module subsystem also allows to load modules 'on-demand'. When an application try to access or configure a network interface, the corresponding module is loaded. I personally prefer to use the 'on-demand' feature of the module subsystem, as this allow you to not have to specify a static list of modules that need to be loaded, and only modules really needed are loaded which saves kernel memory. You can also choose which module to load when there are multiple modules available that support your hardware (which happens quite often). With kernel 2.6.X the module subsystem is configured in the file /etc/modprobe.conf or files in the directory /etc/modprobe.d/. To configure 'on-demand' module loading, on my test system I need to add to the following lines to the configuration : --------- /etc/modprobe.d/local or /etc/modprobe.conf ------ # HP 100VG J2585B PCI card alias eth2 hp100 # AMD AMD PCnet LANCE PCI card alias eth4 pcnet32 # Old AT&T Wavelan ISA card alias eth3 wavelan options wavelan io=0x390 irq=15 ------------------------------------------------------------ Your distribution may already have lines for your interfaces, either replace these or make sure they are correct (some distributions are notorious for picking the wrong driver name in some cases). This file also contains configuration for lot of other subsystems, obviously you don't want to touch that. In this file, you put the name you would like the interface to have (we'll fix that in a minute). Note that for modern PCI cards this is much more straightforward than for old ISA cards. 3) Installing 'ifrename' ------------------------ You will need to install ifrename on your system. 'ifrename' is part of the Wireless Tools package (version 27 and later) and is a complete rewrite of the now obsolete 'nameif'. Some distributions, such as Debian Sarge, offer a separate package for 'ifrename', and in this case you should just install this package. Other distributions may include ifrename as part of their 'wireless-tools' package (this should be the case for Gentoo, Fedora and Mandrake). Other distributions, such as Debian 3.0, don't include ifrename at all, so you should compile yourself a recent version of Wireless Tools (v27 or later) and install it. In any case, you should verify that 'ifrename' is properly installed and check the path needed to call it : -------------------------- > which ifrename /sbin/ifrename -------------------------- Most distributions will install 'ifrename' in '/sbin', while if you compile your own wireless tools, it will be in '/usr/local/sbin'. 4) Making the boot scripts call 'ifrename' ------------------------------------------ You need to make sure 'ifrename' is run at boot time. Most distributions don't do that yet by default. This is a part that is distribution-specific, so you will need to look into your own init files, or ask people familiar with your distribution. It will need to run just before the call to 'ifup' or 'ifconfig' command. In Debian 3.0 and Debian Sarge, it needs to be run from /etc/init.d/networking, which is not the default. The necessary patch is below : ---------------------------------------------------------------- --- networking-orig Wed Feb 18 13:56:23 2004 +++ networking Fri Feb 20 14:51:06 2004 @@ -120,6 +120,15 @@ case "$1" in doopt syncookies no doopt ip_forward no + # Optionally remap interface names based on MAC address. + # '/sbin/ifrename' is part of wireless-tools package. + # /etc/iftab is currently not created by default. Jean II + if [ -x /sbin/ifrename ] && [ -r /etc/iftab ]; then + echo -n "Remapping network interfaces name: " + ifrename -p + echo "done." + fi + echo -n "Configuring network interfaces: " ifup -a echo "done." ---------------------------------------------------------------- Don't forget to set the appropriate path to the ifrename command (see step (3) above). You may also want to also set the proper options for ifrename (check the man page). The option '-p' enables module autoloading compatibility. The default version of 'ifrename' also includes some special Debian support : using "ifrename -p -d", only the proper modules are loaded. If you are using Debian, you should use this option. 5) Renaming interfaces ---------------------- As stated above, we use 'ifrename' to assign names to interfaces. First, you need to get the MAC address of each of your interfaces. You can read the MAC address on the label of the card, or display it using the 'ifconfig -a' command. Remember that the interface won't load yet with the proper name, so you may need to do a bit looking around : ----------------------------- # modprobe pcnet32 # ifconfig eth0 eth0 Link encap:Ethernet HWaddr 00:10:83:34:BA:E5 [...] ----------------------------- The configuration of 'ifrename' is simple, you just specify which name should be used for each MAC address in the file /etc/iftab : --------- /etc/iftab ------------------------ # HP 100VG J2585B PCI card eth2 mac 08:00:09:* # Old AT&T Wavelan ISA card eth3 mac 08:00:0E:* # AMD AMD PCnet LANCE PCI card eth4 mac 00:10:83:* --------------------------------------------- The '*' in the MAC address is a wildcard and allows me to replicate my configuration between multiple identical computers. If you have to manage large number of computers (like a rack of servers or clusters), then you may want to look at other selectors offered by 'ifrename'. To test that ifrename works, do the following : o Load all your drivers, see section (2) o Check /proc/net/dev to see which interface exist o Bring all interfaces down : ifconfig ethX down o Run ifrename o Check each interface with ifconfig o Bring all interfaces up : ifconfig ethX up 6) Configuring interfaces ------------------------- Most likely, your distribution is already doing this part properly. Just assign the proper IP and wireless configuration to each of the interface names you have chosen. This part is distribution specific, and I already document it in the file DISTRIBUTIONS.txt. In Debian, you would need to modify the file /etc/network/interfaces so that it looks something like this : --------- /etc/network/interfaces ----------- # AMD PCnet LANCE PCI card auto eth4 iface eth4 inet dhcp # HP 100VG J2585B PCI card auto eth2 iface eth2 inet static address 10.0.0.2 netmask 255.255.255.0 broadcast 10.0.0.255 gateway 10.0.0.1 --------------------------------------------- This was the last part. Now, at your next boot, all your interfaces should be assigned the proper name and the proper configuration. CONFIGURATION VIA HOTPLUG ------------------------- Dealing with removable interfaces is similar to dealing with built-in interfaces, the main difference is that we will use the Hotplug scripts instead of the init scripts. Another difference is that it will likely require more work on your part because most distributions are not fully ready for it. 1) Applicability ---------------- The Hotplug configuration method is the best choice for any removable network interface, such as : o Pcmcia (16 bits) network cards o CardBus (32 bits) network cards o USB network dongles o Hot-PCI network cards It may also be used to manage other types of network interfaces, although it may not be the best choice for them. 2) How Hotplug works -------------------- Conceptually, Hotplug is very simple. When something interesting happens, the Linux kernel generates an Hotplug event. This runs the proper script from the /etc/hotplug directory. There are 3 types of Hotplug events we care about : o PCI event : a CardBus device is added or removed from the system. The script /etc/hotplug/pci.agent is run. o USB event : a USB device is added or removed from the system. The script /etc/hotplug/usb.agent is run. o Network event : a network interface is added or removed from the system. The script /etc/hotplug/net.agent is run. If we insert a CardBus network card in the system, the following happens : 1) Kernel detects new CardBus device 2) Kernel generates PCI Hotplug event 3) /etc/hotplug/pci.agent runs, finds proper driver module 4) /etc/hotplug/pci.agent loads driver module 5) Driver module initialises, creates new network device 6) Kernel detects new network device 7) Kernel generates Network Hotplug event 8) /etc/hotplug/net.agent runs, configures network device The sequence of events is similar for removals and USB devices. 3) Make sure ifup does not deadlock ----------------------------------- The first problem is that we need to make sure the command 'ifup' does not deadlock by calling itself re-entrantly. If the system has built-in interfaces, the 'ifup' may reenter itself at boot time via Hotplug : 1) Init scripts start running 2) Init script calls 'ifup -a' to initialise built-in network interfaces 3) 'ifup' auto-loads driver module for built-in network interface 'eth4' 4) Driver module initialises, creates new network device 5) Kernel generates Network hotplug event 6) /etc/hotplug/net.agent runs, call 'ifup eth4' Note that you can produce the same reentrancy if you call ifup manually on an interface which module is not yet loaded. The default version of 'ifup' for Debian 3.0 and Debian Sarge is not reentrant and can therefore deadlock if not used properly. The patch to make 'ifup' properly reentrant is available here : http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=231197 Contemporary versions of Debian (Sarge and later) have a net.agent script that contains workarounds to prevents deadlock situations, so for normal use the default 'ifup' should work fine. Other distributions have very different ifup programs and I have not tried those (tell me about it !). 4) Installing Hotplug for Debian Sarge (testing/unstable) --------------------------------------------------------- Thanks to the great work of many people, Debian Sarge has all the necessary packages and hotplug support, and will work mostly 'out of the box'. You will need to install the following packages : o hotplug o ifrename While the installation of Hotplug is simple, its configuration may seem complex. The current network Hotplug script has 3 modes : 'all', 'auto' and 'hotplug'. However for our purpose they all produce the same results when configured. This mode is controlled by the variable NET_AGENT_POLICY in /etc/default/hotplug. In the mode "all", Hotplug will run ifup for all network events. This will result in failure messages if some interfaces have already been configured by the init scripts. This mode is not recommended. In the mode "auto", Hotplug will run ifup only for those interfaces listed in a auto stanza in /etc/network/interfaces. If you choose this mode, you need to put in /etc/network/interfaces a "auto" line for the interfaces you want to control with hotplug. --------- /etc/network/interfaces ----------- # Enable Hotplug support for "auto" mode (Sarge and later) auto eth0 eth1 eth2 eth3 eth4 wlan0 wlan1 prism0 prism1 airo0 airo1 --------------------------------------------- This will result in some failure message at boot time, the init script will attempt to enable all those interfaces, and generate an error for all those not available at this time. It will also generate an error messages for interface which have already been configured by the init scripts. This mode is also not recommended. In the mode "hotplug", hotplug network events are ignored by ifup by default. To enable them you will need to add the following lines to /etc/network/interfaces : --------- /etc/network/interfaces ----------- # Enable Hotplug support for "hotplug" mode (Sarge and later) mapping hotplug script echo --------------------------------------------- To enable them for only selected interfaces, e.g., ethA, make /etc/network/interfaces look like this : --------- /etc/network/interfaces ----------- # Enable Hotplug support for "hotplug" mode (Sarge and later) mapping hotplug script grep map ethA --------------------------------------------- 5) Installing Hotplug for Debian 3.0 ------------------------------------ Debian 3.0 doesn't come by default with hotplug, but the hotplug package is available as regular Debian package (on the CD or downloadable in Debian archive), so you can just install that. Unfortunately, this version of hotplug is not fully compatible with kernel 2.6.X. You will need to do the following modifications to the file /etc/hotplug/net.agent. ------- /etc/hotplug/net.agent ------------------ --- net.agent-d1 Fri Feb 20 18:18:05 2004 +++ net.agent Fri Feb 20 18:22:50 2004 @@ -26,7 +26,7 @@ if [ "$INTERFACE" = "" ]; then fi case $ACTION in -register) +add|register) case $INTERFACE in # interfaces that are registered after being "up" (?) @@ -52,7 +52,7 @@ register) mesg $1 $ACTION event not handled ;; -unregister) +remove|unregister) # Assume that we want to run ifdown no matter what, # because it is not going to remove the data from the # ifstate database otherwise. ------------------------------------------------- Compared to the version in Sarge, this older version of hotplug is much more basic, and doesn't have any scanning at boot time and doesn't need to be enabled in /etc/network/interfaces. 6) Installing hotplug on other distributions -------------------------------------------- The canonical version of hotplug is available at : http://linux-hotplug.sourceforge.net/ Most distributions have customized hotplug packages and chances are that the canonical version won't completely work on your system. All these various changing versions make it difficult for me to tell what exactly needs to be changed in the hotplug scripts to make them work. However, most should work out of the box. My guess is that in a few releases, all these problems will sort themselves out. Just be patient. 7) Dealing with 'init' hotplug ------------------------------ In addition to the standard kernel Hotplug events, modern versions of the Hotplug scripts add init scripts that scan the system buses and generate pseudo Hotplug events at boot time. For the PCI buses, the script /etc/hotplug/pci.rc is run, for the USB bus, /etc/hotplug/usb.rc is run. The end result is that the Hotplug subsystem will also attempt to configure built-in devices : 1) Kernel boots 2) Init runs, start to initialise the OS 3) /etc/hotplug/pci.rc runs, generates pseudo Hotplug event 4) /etc/hotplug/pci.agent loads driver module 5) Driver module initialises, creates new network device 6) Kernel generates Network Hotplug event 7) /etc/hotplug/net.agent runs, configures network device At this point, you realise that at initialisation, both Hotplug and the regular init scripts (see "CONFIGURATION FROM INIT SCRIPTS") are trying to configure the same devices in parallel. This may create problems and is totally redundant. Another reason I don't like this mechanism is that it blindly attempts to load drivers for all hardware present on the system and doesn't use the module loader configuration files to select preferred drivers. It's fairly common to have multiple drivers for a given hardware, and because of Murphy's law, Hotplug will usually load the wrong one. It's also fairly common to have hardware on the system that doesn't need enabling (for example, the IDE controller on my SCSI machine), not loading the driver makes your kernel smaller and boot faster. Hotplug does have a way of disabling the loading of drivers on a case by case basis. Drivers listed in /etc/hotplug/blacklist will not be loaded. Hotplug can be disabled for a whole subsystem by editing the appropriate .rc script in /etc/hotplug, or just deleting/renaming those files. 8) Making hotplug scripts call ifrename --------------------------------------- The last hotplug step is to make sure that 'ifrename' is run by the hotplug subsystem at the right time. As before, we want to run it just before calling 'ifup'. The latest version of the hotplug scripts have this feature integrated. However, you need to check that the path used for calling 'ifrename' is the proper one on your system. And, for older versions of hotplug scripts, you will need to add this support yourself. Check the path for ifrename : -------------------------- > which ifrename /sbin/ifrename -------------------------- The patch to add 'ifrename' to hotplug looks like : ------- /etc/hotplug/net.agent ------------------ --- net.agent-s2 Fri Feb 20 17:18:46 2004 +++ net.agent Fri Feb 20 17:32:43 2004 @@ -40,6 +40,21 @@ add|register) # we can't do much here without distro-specific knowledge # such as whether/how to invoke DHCP, set up bridging, etc. + # Run ifrename as needed - Jean II + # Remap interface names based on MAC address. This works around + # the dreaded configuration problem "all my cards are 'eth0'"... + # This needs to be done before ifup, otherwise ifup will get + # confused by the name change and because iface needs to be + # down to change its name. + if [ -x /sbin/ifrename ] && [ -r /etc/iftab ]; then + debug_mesg invoke ifrename for $INTERFACE + NEWNAME=`/sbin/ifrename -i $INTERFACE` + if [ -n "$NEWNAME" ]; then + debug_mesg iface $INTERFACE is remapped to $NEWNAME + INTERFACE=$NEWNAME + fi; + fi + # RedHat and similar export IN_HOTPLUG=1 if [ -x /sbin/ifup ]; then ------------------------------------------------- If your hotplug scripts already include ifrename support then you should find a section in /etc/hotplug/net.agent looking like the patch above. Otherwise, just cut'n'paste the patch above in the right place. The path for 'ifrename' is used twice above, so don't forget to modify both occurences. 9) Loading driver modules ------------------------- Wow ! The most difficult part is done. In theory, you don't need to do any specific configuration for the driver modules to be loaded. The 'pci.agent' and 'usb.agent' should load the right driver module for you. Also, you don't need to define aliases in /etc/modprobe.d/* or in /etc/modprobe.conf, it's useless and may be counterproductive. If you use a driver compiled statically in the kernel, you also have nothing to do. 10) Renaming interfaces ----------------------- We still use ifrename to assign names to interfaces. The configuration of 'ifrename' is the same. To keep the possibility of having multiple wireless cards (one in each CardBus slot), we use wildcards in both the MAC address and the name : --------- /etc/iftab ----------------------- # SMC 2835W wireless CardBus card prism* mac 00:30:B4:* --------------------------------------------- If you insert two cards, they would be named prism0 and prism1. Note that 'name wildcarding' is a feature only available in 2.6.X and 2.4.30 and later, so if you use older version of 2.4.X you will need to be explicit and list each card separatly : --------- /etc/iftab ----------------------- # SMC 2835W wireless CardBus card prism0 mac 00:30:B4:64:27:8B prism1 mac 00:30:B4:64:27:8D --------------------------------------------- 11) Configuring interfaces ------------------------- At this point, configuration of Hotplug interfaces is done just like their built-in counterparts. This part is still distribution specific, and still already documented in the file DISTRIBUTIONS.txt. In Debian, you would need to modify the file /etc/network/interfaces like this : --------- /etc/network/interfaces ----------- # Enable Hotplug support (Sarge and later) mapping hotplug script grep map prism0 # SMC 2835W wireless CardBus card iface prism0 inet static address 10.0.1.2 netmask 255.255.255.0 broadcast 10.0.1.255 wireless-essid THE_ESSID wireless-mode ad-hoc wireless-channel 5 --------------------------------------------- Note that you should not have wireless-* lines if you are using waproamd to set these parameters. Now, just cross your fingers and plug the card in the slot... PCMCIA INTERFACES (16 bits) --------------------------- The Pcmcia subsystem has quite some legacy, and can use various configuration procedures. The Pcmcia subsystem exclusively uses hotplug for 32 bits cards (if you are using the kernel Pcmcia modules, which is the only option for 2.6.X). For 16 bit cards cardmgr is still required for managing the sockets and loading modules. Cardmgr is configured by files in the /etc/pcmcia directory. To use Hotplug network configuration with 16 bits Pcmcia cards, first make sure the Pcmcia subsystem is properly configured and that cardmgr loads the right driver module (in most case, it should). Then, make sure that you don't have any configuration entries in /etc/pcmcia/network.opts and /etc/pcmcia/wireless.opts. Make sure that none of the entries in your system network configuration use 'eth0' or 'wlan0' (in /etc/network/interfaces for Debian users). Then, just follow the procedure described above for "Configuration Using Hotplug" to configure your network cards. You might want a little bit of explanation on why this magic will work (which would help in case it doesn't work). There are two types of Pcmcia network configuration scripts, available as /etc/pcmcia/network. The original Pcmcia script configures network cards using options found in /etc/pcmcia/network.opts and /etc/pcmcia/wireless.opts. Most distributions replace it with a script calling 'ifup'. By making sure that network.opts and wireless.opts are "empty", we neutralise the first set of scripts. By making sure no system configuration uses 'eth0' or 'wlan0', we neutralise the second set of scripts, the script would call 'ifup' with the default interface name, which is usually 'eth0', ifup would not find a configuration for it and would just ignore it. The card would still be configured because hotplug network events are generated for every interfaces, not only for devices managed by hotplug. So, net.agent would receive an event and perform the necessary steps to configure it. Personally, I'm still using the original Pcmcia scripts for my Pcmcia cards as described in the file PCMCIA.txt, because it still works and I will migrate my complex configurations over time. You can also decide to not use Hotplug for Pcmcia cards and modify the distribution Pcmcia scripts in /etc/pcmcia/* to handle Pcmcia cards with ifrename. You would need to modify /etc/pcmcia/network to add 'ifrename' before 'ifup' the same way it was done for /etc/hotplug/net.agent. But, as in the long term Pcmcia will migrate to Hotplug, I would not bother... MANUAL LOADING, DOCKING STATIONS -------------------------------- Manual loading is used for built-in network interfaces that are only use at specific time, and that you want disabled the rest of the time. We assume that you still use modules so that when the interface is not used you can remove the driver from the kernel. First, you need to set the configuration for those interfaces, the same way it's done for other network interfaces. The main difference is that you need to specify that those interfaces should not be enabled at boot time. It's also a good idea to disable Hotplug init scripts. With Debian, you just need to make sure that the 'auto" keyword doesn't apply to this interface. If you use drivers statically built in the kernel, make sure that ifrename runs at boot time (see CONFIGURATION FROM INIT SCRIPTS). Once it's done, you can just enable and disable those interfaces with 'ifup ethX' and 'ifdown ethX'. If you use both a modular system, make sure that the 'on-demand' module loading is properly configured : --------- /etc/modprobe.d/local or /etc/modprobe.conf ------ # HP 100VG J2585B PCI card alias eth2 hp100 # AMD AMD PCnet LANCE PCI card alias eth4 pcnet32 ------------------------------------------------------------ Then, you should instruct 'ifup' to load module and use ifrename prior to configuring the interface, and remove the module when going down. With Debian, this is done with : --------- /etc/network/interfaces ----------- # AMD AMD PCnet LANCE PCI card # noauto iface eth4 inet dhcp pre-up /sbin/ifrename -p -n eth4 post-down /sbin/modprobe -r eth4 # HP 100VG J2585B PCI card # noauto iface eth2 inet static address 10.0.0.2 netmask 255.255.255.0 broadcast 10.0.0.255 gateway 10.0.0.1 pre-up /sbin/ifrename -p -n eth2 post-down /sbin/modprobe -r eth2 --------------------------------------------- We use the '-n' option of ifrename to specify the name of the interface after renaming. This assume that the mapping for those interfaces don't use wildcards. The '-p' option make sure ifrename probes the module prior to using it. Using "modprobe -r" make sure that if the driver is composed of multiple module all the modules are unloaded. To enable the interface, just use : ----------------------------------- ifup eth4 ----------------------------------- And to disable the interface : ----------------------------------- ifdown eth4 ----------------------------------- This solution is obviously Debian specific, but could be adapted to other distributions. If you can't manage to get your distributions to use those tricks, you can do things manually. If you don't use Hotplug, you enable an interface with : ----------------------------------- modprobe eth4 ifrename ifup eth4 ----------------------------------- If you use hotplug, you only need to do : ----------------------------------- modprobe eth4 ----------------------------------- On the other hand, disabling the interface is done with : ----------------------------------- ifdown eth4 modprobe -r eth4 ----------------------------------- Docking stations for laptops may contain built-in interfaces. My previous laptop had one, and Linux had no support for it. After docking, I was able to bring up the network ISA card in the docking station. However, with most laptops and version of Linux, the issue is that after docking, the new devices are not seen. The solutions is to force a rescan of the PCI bus. Documentation is unclear on that, maybe 'scanpci' may help. To be able to simply manage my docking station, I had created two little scripts to enable and disable my network interface. After docking, you would run : -------- /sbin/dock ---------------------------- #!/bin/sh modprobe eth4 ifrename ifup eth4 ------------------------------------------------ And prior to undocking, you would run : -------- /sbin/undock ---------------------------- #!/bin/sh ifdown eth4 modprobe -r eth4 ------------------------------------------------ Thanks to 'ifrename', the network interface in your dock will always be properly configured regardless of if you have a Pcmcia network card in the Pcmcia slot or not. SCHEMES (MULTI-CONFIG) ---------------------- Most Ethernet cards will only connect to a single network, or can use DHCP to be auto-configured. With Wireless Cards, it's much more likely that you will need multiple configurations, for example at work, at home and on-the-go. Most distributions have various level of support for such schemes. Some distributions offer simple network schemes, while other offer "overall" schemes changing the whole configuration. I document the support for schemes in various distributions in the file DISTRIBUTIONS.txt. You can also use tools such as ifplugd, waproamd or wlandetect. Those tools are a kind of "wireless-DHCP", they attempt to automatically detect the proper wireless configuration and apply it. Most will also attempt to detect network changes. The main limitation of those tools is that they offer very little manual control. If two valid alternatives are possible, you can't switch between them. If a configuration can't be detected, they usually fail. That's the same concept as using DHCP versus Static IP addresses. Some people are very happy with DHCP, my style is Static IP addresses. If you use Debian and want to use simple manual schemes, these are the things you need to do. 1) Make sure that 'ifscheme' and 'ifscheme-mapping' are installed on the system. You may find them in a separate tar file on my web site. 2) Check the path for 'ifscheme-mapping' (using whereis). 3) Modify you /etc/network/interface to add proper mapping and configuration. ------- /etc/network/interfaces ---------------------- # Enable Hotplug support (Sarge and later) mapping hotplug script echo # SMC 2835W wireless CardBus card mapping prism0 script /sbin/ifscheme-mapping iface prism0-any inet dhcp wireless-essid any wireless-mode managed iface prism0-adhoc inet static address 10.0.1.2 network 10.0.1.0 netmask 255.255.255.0 broadcast 10.0.1.255 wireless-essid THE_ESSID wireless-mode ad-hoc wireless-channel 5 iface prism0-other inet static address 10.10.10.2 network 10.10.10.0 netmask 255.255.255.0 broadcast 10.10.10.255 wireless-essid ANOTHER_ESSID wireless-mode ad-hoc wireless-key "s:secure" ------------------------------------------------------ FIRMWARE LOADING ---------------- A lot of modern wireless card don't have built in firmware and need firmware loading. Recent kernels (2.6.X) have a firmware loader. These are a few notes on how to use it. First, read the documentation coming with your driver, because each driver has specificities (like the name of the firmware file it requires). Some drivers may offer additional ways to load the firmware, but in the long term things should be standardised around the hotplug method to simplify packaging in distributions. You need to compile your kernel with firmware loading (CONFIG_FW_LOADER in "Generic Driver Options"). If your driver was built from the kernel, chances are that it enabled this feature already. Make sure you boot from this new kernel. The 'sysfs' file system must be mounted. The easiest is to mount it at boot time, add a line for it in /etc/fstab : -------- /etc/fstab ------------------------------ sysfs /sys sysfs defaults 0 0 -------------------------------------------------- Then, you add the firmware file in the directory where it's expected, which is /usr/lib/hotplug/firmware/ in most cases. Most distributions nowadays have a version of the Hotplug scripts that knows how to deal with firmware. If it is not the case, just grab the 'firmware.agent' file from an alternate source and copy it into your /etc/hotplug directory (make sure it's executable). You can try the canonical version : http://linux-hotplug.sourceforge.net/ Or Debian's version : http://packages.debian.org/unstable/admin/hotplug Note that firmware loading will usually only work with interfaces that are fully managed by Hotplug. This is the only way to ensure the that proper sequence of action is happening in the right order every time. Firmware loading may not work properly for interfaces configured in the init scripts. This means that if you have a built-in interface that require firmware loading, you should just use manage those interfaces like removable interfaces (see section above). However, interface configuration need to be explicitly triggered at boot time. One possibility is to set-up Hotplug to be run from the init script at boot time. This is usually an option for recent distributions (it's not the case for Hotplug in Debian 3.0). But, we have seen that this has some issues. The other possibility is to use an hybrid between the init script method and the hotplug method. First, you need to add an alias for the driver in /etc/modprobe.conf. Then, you need to specify a mapping for this interface in /etc/iftab, and specify a configuration for this interface and that it is enabled at boot time. Lastly, you make sure that the network init scripts run 'ifrename -p'. 'ifrename' will trigger the module to load, and all the Hotplug events will be generated properly to configure the interface. DEVICES WITH MULTIPLE NAMES --------------------------- Some wireless drivers offer multiple network interfaces for the same device. A classical example is the Aironet driver that creates a 'ethX' and 'wifiY' for each card. 'ifrename' allows you a finer selection of interfaces than 'nameif'. For example, to only rename the pseudo-Ethernet network interface name of the Aironet driver, you would do : --------- /etc/iftab ----------------------- # Cisco Aironet 350 wireless Pcmcia card airo* mac 00:07:0E:* arp 1 --------------------------------------------- After that, your device would be available through 'eth0' and 'wifi0'. You can rename both interfaces. You just need to remember that 'ifrename' starts matching from the last line of the file, so you would do : --------- /etc/iftab ----------------------- # Cisco Aironet 350 wireless Pcmcia card wifi* mac 00:07:0E:* airo* mac 00:07:0E:* arp 1 --------------------------------------------- The current version of 'ifrename' supports only the most useful selectors, but it is architectured such as adding selectors is relatively trivial. If you find a case that 'ifrename' can't handle, you should just extend it. DEVICES WITHOUT MAC ADDRESSES ----------------------------- Most Ethernet and Wireless devices have a fixed and unique MAC address, and it is therefore advised to name them based on this criteria. However, there are also network interfaces that don't have a fixed and unique MAC address, for example Ethernet over USB, IP over FireWire, PPP and tunnel interfaces. The driver for those devices creates the interface with a name specific to the driver, such as ppp* for PPP interfaces and usb* for Ethernet over USB, and therefore they are easy to identify and configure, and few users feel the need to rename them. Moreover, some of them, such as PPP, have their own configuration scripts and methodology addressing their unique needs. There are a few cases where you might want to rename interfaces without MAC addresses. One example is two Ethernet over USB dongles. The way to do this is to use alternate ifrename selectors. Choosing the right selector depends on what you want to achieve. A quick theoretical example to illustrate : --------- /etc/iftab ----------------------- # All other usbnet devices usb* driver usbnet # Specific usbnet devices usb-p firmware "Prolific PL-2301/PL-2302" usb-4 bus-info usb-00:02.0-1.4 --------------------------------------------- TROUBLESHOOTING --------------- If your interface doesn't show up as expected with ifconfig, you will need to find out why. First, you need to be familiar with the sequence of actions in the system and find which one did not happen. You need to check that the driver module(s) was loaded using 'lsmod'. You need to check if the interface was properly renamed with 'ifrename'. You can use 'ifrename -D -V' to debug your /etc/iftab. Get the list of interfaces on your system with 'ifconfig -a' or 'cat /proc/net/dev', and check if an interface is using the name you assigned or 'eth0'. Check any suspicious interfaces with 'ifconfig eth0', and check its MAC address. Note that some rare drivers don't have a proper MAC address before brought up, which fools ifrename. Verify that no line in /etc/iftab matches the all-zero MAC address. The all-zero MAC address matches the loopback interface 'lo' and various pseudo network devices, renaming the loopback interface is highly discouraged. You need to check which configuration was given to the interface using 'ifconfig' and 'iwconfig'. The Hotplug subsystem has also good debugging facilities. To enable Hotplug debugging, just make sure the variable DEBUG is defined in /sbin/hotplug : --------- /sbin/hotplug ------------------------------ --- /sbin/hotplug-old Tue Mar 26 09:00:20 2002 +++ /sbin/hotplug Fri Feb 20 18:40:38 2004 @@ -22,7 +22,7 @@ cd /etc/hotplug . hotplug.functions -# DEBUG=yes export DEBUG +DEBUG=yes export DEBUG if [ "$DEBUG" != "" ]; then mesg "arguments ($*) env (`env`)" ------------------------------------------------------ Then, you can check your message logs for Hotplug events with 'tail -f /var/log/messages'. Verify that the various Hotplug events happen as expected (pci, firmware, net...), and verify the log messages from 'net.agent'. Have fun... Jean