Re: [PATCH] RDMA/core: use rdma_cap_iw_cm() in rdma_resolve_route()

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Sep 20, 2023 at 08:44:52AM +0300, Leon Romanovsky wrote:
> On Tue, Sep 19, 2023 at 04:08:38PM -0400, Vitaly Mayatskikh wrote:
> > On Tue, Sep 19, 2023 at 4:06 PM Vitaly Mayatskikh <vitaly@enfabrica.net> wrote:
> > >
> > > On Tue, Sep 19, 2023 at 3:21 AM Leon Romanovsky <leon@kernel.org> wrote:
> > >
> > > > I see that rdma_protocol_iwarp() is used in other places in cma.c too,
> > > > Don't they need to be updated too?
> > > >
> > > > Also I see that we have check for protocol RoCE in else before the
> > > > changed line, shouldn't all cma.c be changed to rdma_cap_*_cm() calls?
> > > >
> > > >   3376         else if (rdma_protocol_roce(id->device, id->port_num)) {
> > >
> > > I can't really judge, but looking around in the code it seems that
> > > some if not all of
> > > those cma.c functions that are checking for the protocol - they only
> > > called from the
> > > drivers that actually use the protocol. For example, iSER.
> 
> Just to make sure that we are using correct terminology - iSER is ULP
> (upper layer protocol) and not driver. 
> 
> > >
> > > Our driver does not support iWarp, but implements IW_CM callbacks. The patch has
> > > the only fix that was needed to make it work w/o a full blown iWarp.
> 
> It is hard to say without having driver in-tree and seeing the result of
> ib_device_check_mandatory() in regards of kverbs_provider variable.
> 
> Does any existing in-tree driver require the proposed change in
> rdma-cm?

Yes, lets see a driver first please.

iWarp CM is tightly tied to iWarp, I have a hard time understanding
how you could have the CM without supporting iWarp too.

Jason