[ofa-general] [PATCH RFC] rds: add iwarp support

[Or Gerlitz]

On 7/4/08, Jon Mason <jon at opengridcomputing.com> wrote:

>  The bulk of the code is a copy of the RDS IB with the changes necessary
>  to get iWARP working.

So the intension is to have two code bases "RDS/IB" and "RDS/iWARP"?

if yes, why? the rds code was enhanced to support credit management
and my understanding is that this paves the way to have joint code, as
done for any other RDMA ULP that was ported to iWARP (rNFS, Open-MPI,
Lustre, etc)

>  The logic in the RDS connection setup to determine which method to use,
>  requires a new function in the core IB logic to translate a given IP
>  Address to the IB/iWARP device associated with it.  The only place this
>  can be determined is in the pre-existing structs in cma.c.

.laddr_check is not used in this patchset, and I could not understand
its role from this description.

>  Olaf, if you would like to pull this into your personal tree, I can
>  provide you with delta patches until it is accepted into the OFED 1.4
>  tree (assuming that you have no major problem with the code below).

As its an RFC, lets have it go a little review before pulling it,
unless the intension is to put it in an experimental branch and let
people play with. Anyway, structured set of delta patches is always
prefered, for review, acceptance, debugging and maintainace reasons.

>  --- a/drivers/infiniband/core/cma.c
>  +++ b/drivers/infiniband/core/cma.c

>  +struct ib_device *ipaddr_to_ibdev(u32 addr)

is there a reason why rdma_resolve_addr(NULL, addr, timeout) wouldn't
do what ipaddr_to_ibdev does?

Also, from quick looking I understand that this patchset adds some
sort of fmr pool for the new fmrs within the rds code, can this pool
be generalized and added to the core module such that other ULPs would
be able to enjoy it?

Or.