RE: Question on IPv6 over IB

Vivek & Roland,

It will be much easier to integrate into the Linux implementation if the pad
is after the MAC address of the net-device in the Source/Target Link-layer
address option as the net-device's mac address is directly used in
generating the IPv4 ARP packets and the Source Link-layer address option.

Thanks,
Anand.
-----Original Message-----
From: Roland Dreier [mailto:[email protected]] 
Sent: Thursday, January 13, 2005 4:33 PM
To: Vivek Kashyap
Cc: Anand Parthasarathy; [email protected]
Subject: Re: [Ipoverib] Question on IPv6 over IB

    Vivek> Isn't the length of the option already 8 when used on
    Vivek> Ethernet (type+length+ 48bit)? Then it wouldn't require the
    Vivek> padding.

Of course you are correct.  The 6 octet length of the ethernet address
confused me, sorry.

    Vivek> Will it not align the local address at 2 byte boundary as
    Vivek> against 4?

That's true.  It seems there is some precedent for putting the padding at
both ends: IPv6-over-ARCnet (RFC 2497) puts the (one octet!?) ARCnet address
at the beginning of the option followed by 5 octects of padding, while
IPv6-over-Fibre Channel (RFC 3381) puts two octets of padding at the
beginning followed by 12 octets of address.

I will generate a patch for the Linux kernel to pad appropriately and
forward it to the networking maintainers.

Does anyone know what Solaris's IPoIB implementation does?

Thanks,
  Roland