svn commit: r273806 - in head/contrib/ofed: libcxgb4 libcxgb4/src usr.lib usr.lib/libcxgb4

[Edward Tomasz Napierała]

On 1029T1123, Navdeep Parhar wrote:
> On Wed, Oct 29, 2014 at 10:56:04AM +0100, Edward Tomasz Napierała wrote:
> > On 1029T0115, Navdeep Parhar wrote:
> > > Author: np
> > > Date: Wed Oct 29 01:15:48 2014
> > > New Revision: 273806
> > > URL: https://svnweb.freebsd.org/changeset/base/273806
> > > 
> > > Log:
> > >   Userspace library for Chelsio's Terminator 5 based iWARP RNICs (pretty
> > >   much every T5 card that does _not_ have "-SO" in its name is RDMA
> > >   capable).
> > 
> > Yay!  This means we could add iSER without using the ICL_PROXY hack.
> > Well, assuming it's possible to "hand off" RDMA connection from userspace
> > to the kernel.  Is it?
> 
> Yes, this should be doable.  The connection is just another TCP endpoint
> tracked like all others in the kernel.

:-)

> By the way, iSER is an unnecessary layer if you're using a T5 NIC.
> It'll work, sure, but you'll run iSER/RDMA/TOE when you could simply run
> iSCSI/TOE with full zero copy everywhere.  Comes out to the same result
> with a much simpler stack.  I think iSER makes sense for gear that does
> RDMA but not iSCSI natively.

True.  IMHO the biggest difference is that iWARP is universal, ie.
the NIC doesn't need to have any iSCSI-specific functionality.  On
the other hand, iSER cannot talk to non-iSER, while iSCSI offload
in your NICs talks ordinary iSCSI on the wire.