-----"Jason Gunthorpe" [off-list ref] wrote: -----

To: "Bernard Metzler" <redacted>
From: "Jason Gunthorpe" <jgg@ziepe.ca>
Date: 02/16/2021 07:12PM
Cc: "Chuck Lever" <redacted>, "linux-rdma"
[off-list ref], "Benjamin Coddington"
[off-list ref]
Subject: [EXTERNAL] Re: directing soft iWARP traffic through a secure
tunnel

On Tue, Feb 16, 2021 at 12:27:33PM +0000, Bernard Metzler wrote:

quoted

rdma_port_get_link_layer(). Asking the RDMA core experts -
would a gid of zero have any side effects or bad implications,
since that ID is by no means unique?

Yeah, it is clearly not ok. Generate a random private GUID?

Right. But that's unfortunately not sufficient to get
devices w/o hardware addresses handled by the CM code.
The net_device remains w/o HW address, and that is where
CM gets the source HW address from (rdma_copy_src_l2_addr()
etc.). I don't have a good idea yet how to get it working.

Actually, all this GID and GUID and friends for iWARP
CM looks more like squeezing things into InfiniBand terms,
where we could just rely on plain ARP and IP 
(ARP resolve interface, see if there is an RDMA device
bound to, done)... or do I miss something?


On another related topic - the RDMA CM should let
the provider know if an IPv6 address is to be bound
AFONLY. This becomes needed in the current NFS/RDMA
over iWarp case, where NFS listeners bind to both IPv4
and IPv6 address with same port number. The code
correctly calls rdma_set_afonly() on the cmid, but
there is no mechanic to pass that information down
to the provider. So the second bind fails.
I'll send a small patch which enables that for
iwcm and siw, hoping for acceptance.

Many thanks,
Bernard.