802.1Q VLAN implementation for Linux

MTU problems exist for many ethernet drivers. Other than that, things seem fairly stable!

PLUG:   Check out my company that makes traffic generation and WAN simulation test equipment based on the Linux operating system:

Let us help you test your DSL, Cable Access, Satellite and other network systems!

Join the vlan mailing list, After that, to post, send mail to vlan@ns1.wanfear.com.

Submit a bug/issue/enhancement with the: VLAN Bugzilla

I hear that the 2.2/2.4 kernel patches have worked with these (and other, I'm sure) systems:

• Cisco: {Catalyst: 6509}, 3Com: {Corebuilder, Netbuilder II, SuperStack II switch 630}, Alpine: {3804(SMMi,F32Ti)} Extreme Ntwks {Summit 48, 48i, 5i} Foundry: {ServerIronXL, FastIron}
• Alteon ACENic Gigabit, 3Com 3c509, realtek RTL8029(AS), RTL8139, DEC DC21140 (tulip), DFE-570TX quad-21143, Intel PRO/1000 with Intel's driver

Performance: The difference in running traffic over VLANs v/s regular ethernet is very slight. If someone has done some sort of benchmark, I'll be happy to place it here!

• VLAN Installation & Configuration info.
• CVS Access.
• VLAN HOWTO/FAQ (Some CISCO & 3COM specific info too.)
• Another VLAN Recipe (Some info specific to Intel EEPRO Nics too.)
• VLAN Research page in Japanese
• VLAN page translated to English
• IEEE 802.1Q Standard

• Implements 802.1Q VLAN spec.
• Implements support for a non-standard (as far as I know) MAC-based VLAN functionality.
• Can support up to 4094 VLANs per ethernet interface.
• Scales well in critical paths: O(n), where n is the number of PHYSICAL ethernet interfaces, and that is only on ingress. O(1) in every other critical path, as far as I know.
• Supports MULTICAST
• Can change MAC address of VLAN.
• Multiple naming conventions supported, and adjustable at runtime.
• Optional header-reordering, to make the VLAN interface look JUST LIKE an Ethernet interface. This fixes some problems with DHCPd and anything else that uses a SOCK_PACKET socket. Default setting is off, which works for every other protocol I know about, and is slightly faster.

• vconfig binary for x86
• vconfig binary for StrongArm

• Release 1.8 (gz)   For Kernel: 2.4.21+   Sept 30, 2003:
  
  • Updated MAC-VLAN code and completed testing. Based on Alex Zeffertt's work but much has been re-written and he cannot be held responsible! Please send all bug reports to the VLAN mailing list. The Candela Technologies unified patch is the thing to apply now, and it contains various other not-necessarily-VLAN related bits and pieces.
• Release 1.7m (gz)   For Kernel: 2.4.14+   Feb 27, 2003:
  
  • Added Alex Zeffertt's MAC-based VLAN code. Not fully functional yet (mostly because I broke his original work...gonna fix it up soon. Grab & use his raw patch* files in the meantime.
• Release 1.7 (gz)   For Kernel: 2.4.14+   Feb 27, 2003:
  
  • Clarified the license for vconfig (GPL). Other small tweaks.
• Release 1.6 (gz)   For Kernel: 2.4.14+   March 24, 2002:
  
  • Removed 2.4 kernel patch: It's in the standard kernel now.
  • Updated vconfig to fix some compile issues, and enable cross-compilation to the StrongARM platform (changes should help other cross-compile attempts too.)
• Release 1.5 (gz)   For Kernel: 2.4.12-pre5   October 22, 2001:
  
  • Mostly added other peoples fixes and patches (thanks folks!)
  • Finally fixed mc-list leakage (Ard van Breemen)
  • Flush mc-list at vlan-destory (Ard van Breemen)
  • Add vconfig man page to distribution (Ard van Breemen)
  • Fix problem with /proc and renaming VLAN devices (af AT devcon D.T net)
  • Add relatively large change by Nick Eggelston that makes VLAN devices more transparent to tools like tcpdump and other raw packet snoopers. This will only be enabled when the REORDER_HDR flag is set.
• Release 1.4 (gz)   For Kernel: 2.4.8   August 16, 2001:
  
  • Code should no longer require /proc interface in order to get at the IOCTLs. The IOCTLs are now tied to sockets. When using modules, it may auto-load now, too...
  • Fixed format string error in proc fs display.
  • Fixed crash bug relating to memory allocation with locks held (we now use GF_ATOMIC)
  • hard_start_xmit will now grow the packet header if there is not enough headroom. This may fix an MPLS-over-VLAN problem, though the real solution is to make MPLS allocate more headroom anyway...
  • vconfig was changed to use the new IOCTL API, and the old vconfig WILL NOT WORK with this or any newer patches...
• Release 1.0.3 (gz)   For Kernel: 2.4.7   August 5, 2001:
  
  • Re-worked code to be more stable and more in-line with what the kernel maintainers want to see before the VLAN patch is included into the kernel.
  • One of those requests was to change the default naming scheme to eth0.5, for a VLAN of VID 5 on eth0. You can over-ride this naming behaviour with the vconfig tool.
  • There were *NO* changes to the 2.2 series patch, and I don't expect to ever make any more changes there...
• Release 1.0.1 (gz)   For Kernel: 2.2.18/19, 2.4.3-pre3   April 16, 2001:
  
  • Incorporated a fix for changing a MAC on a VLAN, it now correctly sets PACKET_HOST. Thanks to Martin Bokaemper for this one.
  • The 2.4 series patch should now compile as a module, thanks to a tweak from someone who's mail I have lost! Anyway, 3 cheers to the un-named coder!
  • There were *NO* changes to the 2.2 series patch, though I did verify that it seems to work fine with the 2.2.19 kernel.
• Release 1.0.0 (gz)   For Kernel: 2.2.18, 2.4.0   Jan 14, 2001:
  
  • Really fixed (and tested) MAC change-ability. When you set the MAC address on a VLAN, it will also attempt to set the underlying device to PROMISCious mode (otherwise, the VLAN will not receive any packets.)
  • Hashed-device lookup is disabled by default because some people had trouble with the 'lo' device. Please feel free to re-enable by editing the line in net/core/dev.c (search for #define BEN_FAST_DEV_LOOKUP).
  • vconfig should warn when creating VLAN 1, because that VLAN is not compatible with many switches.
• Release 0.0.15 (gz)   For Kernel: 2.2.18, 2.4.prerelease   Dec 31, 2000:
  
  • Merged most of Matti Aarnio's patches. This means no significant patch to eth.c now, and will help port VLANs to non-ethernet devices (ie ppp, TokenRing??).
  • Setting the MAC address should work now..I think it was broken before.
  • Miscellaneous code re-organization to make patches to existing files smaller.
• Release 0.0.14 (gz)   For Kernel: 2.2.17, 2.4.pre9   Oct 26, 2000:
  

  This code seems pretty stable.

  • Removed vlan-space-per-machine, so vlan-space-per-NIC is mandatory now.
  • DHCP might work now, as I've added support for encapsulating regular ethernet frames if they are sent to the vlan driver.
  • Fixed up the name/index hashing stuff to handle changing the name on a device.
  • Took out default VID & default priority, as their usefullness was in question, and the code was broken anyway.
• Release 0.0.13 (gz)   For Kernel: 2.2.17, 2.4.pre9   Oct 11, 2000:
  KNOWN TO BE BUSTED, here for posterity's sake.
  • Added support for MULTICAST to the VLAN devices. Thanks to Gleb & Co for most of that code.
  • Added the ability to set the MAC address on the VLAN. For now, you'll either need to set your Ethernet NIC into PROMISC mode, or maybe figure out some multi-cast ethernet address to set on the NIC. This has not been tested well at all.
  • Added a hashed device-name lookup scheme. This greatly speeds up ifconfig -a. I was able to run an ifconfig -a in 20 seconds on a Celeron 500, with 4000 vlan devices configured!!
  • Added vlan_test.pl to help me find dumb bugs. Feel free to make this much more powerful, and send the code back to me!
  • vconfig.c has been converted to C code now, instead of C++. Thanks to MATHIEU.
  • Significantly cleaned up the code w/out decreasing any useful functionality, I believe.
  • Removed the DHCP stuff from the VLAN distribution.
• Release 0.0.12 (gz)   For Kernel: 2.2.16, 2.4.pre7   August 27, 2000:
  Added ability to re-order the VLAN packet so that it looks like a real ethernet packet for the ingress pathway. This should help DHCP and other programs that insist on reading the raw buffer and then make assumptions about byte offsets. I don't have a good way to test this fully, so consider it experimental :) This behavior can be changed at run-time, and is set on a per-VLAN basis. The default is NOT to reorder the header, which has been the only behavior up untill this point. The vconfig program can set/clear the flag, by using a VLAN IOCTL. You can read the flag's value from the /proc/net/vlan/vlan* files.

  You can also set a default priority on a NON-VLAN device. This priority will only be used when the default_VID for the device is set as well. This priority won't be mapped anywhere, just copied straight into the skb->priority. It is a uint16.

  The 2.3 patch is now the 2.4 patch, and it has been tested against 2.4.pre7.

• Release 0.0.11 (gz)   For Kernel: 2.2.13/14, 2.3.99   April 23, 2000:
  Added real support for PRIORITY. Through IOCTL calls (see the vconfig program), you can set explicit ingress and egress mappings to/from the VLAN QOS bits and the sk_buff->priority field. This is not tested very well, as I don't know much about how people really use the priority field... Took out the round-robin aggretation that went in in rls 0.10, as it was mainly just a hack, and doing link aggregation at a lower level and then putting VLAN on top of that virtual device probably makes more sense. The vconfig program changed to support the new features..here's it's new usage:
  

  Usage: add             [interface-name] [vlan_id]
         rem             [vlan-name]
         set_dflt        [interface-name] [vlan_id]
         add_port        [port-name]      [vlan_id]
         rem_port        [port-name]      [vlan_id]
         set_egress_map  [vlan-name]      [skb_priority]   [vlan_qos]
         set_ingress_map [vlan-name]      [skb_priority]   [vlan_qos]
         set_name_type   [name-type]
         set_bind_mode   [bind-type]
  
  * The [interface-name] is the name of the ethernet card that hosts
    the VLAN you are talking about.
  * The port-name is the name of the physical interface that a VLAN
    may be attached to.
  * The vlan_id is the identifier (0-4095) of the VLAN you are operating on.
  * skb_priority is the priority in the socket buffer (sk_buff).
  * vlan_qos is the 3 bit priority in the VLAN header
  * name-type:  VLAN_PLUS_VID (vlan0005), VLAN_PLUS_VID_NO_PAD (vlan5),
                DEV_PLUS_VID (eth0.0005), DEV_PLUS_VID_NO_PAD (eth0.5)
  * bind-type:  PER_DEVICE  # Allows vlan 5 on eth0 and eth1 to be unique.
                PER_KERNEL  # Forces vlan 5 to be unique across all devices.
  

  The 2.3 patches have been ported foward to 2.3.99, thanks to Patrick for the vlanproc.c updates!

1. Install and configure cvs on your machine.
2. Specify the vlan repository:
   export CVSROOT=:pserver:anonymous@ns1.wanfear.com:/home/cvs/vlan
3. Log in to the repository:
   cvs login     (PASSWORD: anonymous)
4. Check out the source:
   mkdir vlan; cd vlan; cvs -z3 checkout vlan

Now, you should have a vlan directory in your home directory. You only have to patch the kernel if you are using Linux 2.4.14 or earlier. Now, read the README or other docs to figure out what kernel it patches against. A list of mirrors are kept at www.kernel.org. Unzip and un-tar this in your home directory as well, which should create a linux directory in your $HOME directory. Example: tar -xvzf linux-2.2.14.tar.gz

Now add the VLAN kernel changes to the kernel if your kernel requires it. I finally figured out how to do patches that diff can handle (I think I did it right at least!). You will find the patch in the vlan directory. It will be called: vlan.patch, or something equally straight-foward. Apply the patch to your kernel:

cd $HOME/linux
patch -p 1 < $HOME/vlan/[vlan.patch]

Your new, patched, kernel should be in your INCLUDE path before trying to compile the vconfig program. One way to get things working is to link $HOME/linux to the 'linux' directory that you just un-zipped and patched. A command might be something like: cd $HOME; ln -s /home/greear/kernel/2.4/linux.dev linux

Build the vconfig program in the $HOME/vlan directory:
cd $HOME/vlan
make

Now, time to compile your new kernel! Use the make xconfig command in your $HOME/linux directory to select your kernel options. The option related to 802.1Q VLANs is found under the Networking options. If the option is not highlighted, make sure you select "Experimental Drivers" in one of the first xconfig menus.

Assuming your kernel compiled cleanly (yell if it didn't and you think my code broke it!!), you are now ready to try it out!! Install your kernel in the normal manner (fix up your /etc/lilo.conf file appropriately and run lilo as root.) Reboot your computer and choose your new kernel.

As your computer comes back to life, there will be little sign that you are now 802.1Q capable, other than a line spit out during the boot process. There should be a config programs in your $HOME/vlan directory: vconfig. vconfig is used to create and destroy VLAN devices. So, lets create a VLAN device on your first ethernet NIC. vconfig<return> will list a short spiel on how to use it. The vconfig command I usually use is:

vconfig add eth0 5

This attempts to create a VLAN device with VLAN-ID of 5 on the eth0 device. If you want to delete a VLAN, use something like:

vconfig rem eth0.5

You will also need to give it an ip, eg: ifconfig -i eth0.5 192.168.2.1
and configure it UP: ifconfig -i eth0.5 up

NOTE: You can get lots of VLAN related configuration information from the /proc/net/vlan/* files by using 'cat' or 'more' to look at them.

Please get in contact with me if you have suggestions, patches, or other comments.