[openib-general] Re: rdma_bind_addr question

[Tom Tucker]

On Thu, 2005-10-20 at 13:09 -0700, Sean Hefty wrote:
> Tom Tucker wrote:
> > BTW -- I think this means that we need an ARPHRD_IWARP type.
> 
> It may be that the CMA can simply look in its local device list for an ib_device 
> that has a given MAC address.

After pondering this, I think you're correct. There is one issue,
however. Currently, the GID is stored beginning at the fourth byte of
the dev_addr for IBoIB, but the Ethernet MAC address begins at byte 0.
Is it possible to move this 4B quantity to follow the GID? If so, we
could pad the dev_addr for iWARP devices with zeroes and use the exact
same code to search the cma device table. If the device is found, it
already has a type in the ib_device structure to distinguish between IB
and iWARP devices.

If the caller gave us an IP address for a dumb Ethernet device, we would
go looking for it in the cma device list and simply not find it. It
would still fail, just later, and would avoid a new ARPHRD type. 

> 
> I don't know the detail of how iWarp will work with this.  Will iWarp need a 
> call similar to ib_translate_addr() to translate an IP address into a MAC 
> address?  Are the MAC addresses stored with the ib_device somehow?
> 
> - Sean