Re: [PATCH net-next v2] net/smc: add full IPv6 support for SMC

[Leon Romanovsky <leon@kernel.org>]

On Tue, Oct 28, 2025 at 05:54:50PM +0800, D. Wythe wrote:
> On Mon, Oct 27, 2025 at 03:42:27PM +0200, Leon Romanovsky wrote:
> > On Wed, Oct 22, 2025 at 11:23:09AM +0800, D. Wythe wrote:
> > > The current SMC implementation is IPv4-centric. While it contains a
> > > workaround for IPv4-mapped IPv6 addresses, it lacks a functional path
> > > for native IPv6, preventing its use in modern dual-stack or IPv6-only
> > > networks.
> > > 
> > > This patch introduces full, native IPv6 support by refactoring the
> > > address handling mechanism to be IP-version agnostic, which is
> > > achieved by:
> > > 
> > > - Introducing a generic `struct smc_ipaddr` to abstract IP addresses.
> > > - Implementing an IPv6-specific route lookup function.
> > > - Extend GID matching logic for both IPv4 and IPv6 addresses
> > > 
> > > With these changes, SMC can now discover RDMA devices and establish
> > > connections over both native IPv4 and IPv6 networks.
> > 
> > Why can't you use rdma-cm in-kernel API like any other in-kernel RDMA consumers?
> > 
> > Thanks
> > 
> > >
> 
> Hi Leon,
> 
> Regarding RDMA-CM, I’m not sure if I’ve fully grasped your point, but
> based on my current understanding, I believe SMC cannot use RDMA-CM.
> There are a few reasons for this:
> 
> Firstly, SMC is designed to work not only with RDMA devices but also
> needs to negotiate with DIBS(DIRECT INTERNAL BUFFER SHARING) devices. This
> means we must support scenarios where no RDMA device is present.
> Therefore, we require a round of out-of-band negotiation regardless of
> the final device choice. In this context, even if we ultimately select
> an RDMA device, using rdma-cm to establish the connection would be
> redundant.

Ahh, yes, I always failed to remember this.

Thanks

> 
> Additionally, SMC requires multiplexing multiple connections over a
> single QP. We need to decide during the out-of-band negotiation which
> specific QP to reuse for the connection. From what I know, rdma-cm does
> not seem to offer this capability either.
> 
> Best regards,
> D. Wythe
>