cd drivers/net
find . -name \*.c -or -name \*.h | xargs grep set_mac_address


Well, following my original post, I posted a similar request to the LKML with no response. I then followed up by sending the follwing message to Donald Becker at: vortex -at- scyld.com:

Hi,

I've got this address as a contact address from the top of the latest BK
commit of the 3c59x.c driver in the Linux Kernel.

I'm not 100% sure this is the correct place to post my query, but I've
already posted to the LKML to no luck.  If this is not the right place,
I would greatly appreciate if you could let me know where is the correct
place to post this query...

Basically, I am trying to use two 3c905C (Tornado) NIC's with the 2.4.23
bonding driver.  This driver needs the base NIC driver to support
changing the card MAC address whilst the card is open.  It appears that
the 3c59x Driver does not support this operation.

I have looked at the 3c59x code, and the Bonding driver code, as well as
the code for drivers which do support this operation (such as the Intel
EtherExpress e100) to see if I can glean any useful information. 
Unfortunately, I'm not as familiar with the kernel code as I would like
to be, so I'm a little in the dark.

I have found that cards which do support changing the MAC address of the
card while the device is open have a function binding to the
set_mac_address function in the net_device structure for the card.  What
I can't work out is whether it would be possible to write an appropriate
function for the 3c59x driver.

I've also looked at other versions of the driver at scyld.com, and also
at 3Com's own 3c90x driver to see if I can get any useful information
there - or to see if there is an implementation of the set_mac_address
function that I can adapt to the version of the driver included with the
stock 2.4.23 kernel, also to no luck.

Is it possible to change the MAC address of these cards whilst the
device is open?  If not, is it simply a driver limitation, or is there a
hardware limitation which prevents the change?  Finally, if it is just a
driver limitation, are there any good resources on the kernel drivers
that I can look at that would help me start to write a patch to enable
this functionality?

Any help or advice would be greatly appreciated.  In the meantime, my
searches through Google and other kernel sites continues...

I very promptly received this reply:

On 16 Dec 2003, Mike Insch wrote:

> Basically, I am trying to use two 3c905C (Tornado) NIC's with the 2.4.23
> bonding driver.  This driver needs the base NIC driver to support
> changing the card MAC address whilst the card is open.

This is deeply flawed: there are inherent race and packet drop
problems.
Most NICs do not allow changing the driver MAC address when the receiver
is active because of the problems.

> It appears that the 3c59x Driver does not support this operation.
> I have looked at the 3c59x code, and the Bonding driver code, as well as
> the code for drivers which do support this operation (such as the Intel
> EtherExpress e100) to see if I can glean any useful information. 

The e100 is a good example of why this is silly.  The chip changes the
MAC address using the same command queue as transmitting a packet, which
means that the operation might be delayed for a significant period of
time.  To know when the command is executed, the driver must stop the
transmit queue, wait until the queued packets are sent, send the
'CmdIASetup' update command, and then resume normal operation.  And
while the command is running, the receiver is disabled, so the NIC is
silently dropping packets.  Doh!  It's exactly as bad as down-change-up,
except that the driver is now more complex.

So, It seems at the moment that a software solution to my particular problem is not entirely the right way to go. Time to get managemnt to upgrade our switches!

Thanks to Donald for a fast and extremely informative reply :)