Re: [PATCH rdma-next 0/5] RDMA/mlx5: Add NIC_RX_CLASS steering domain for RDMA classification

[Leon Romanovsky <leon@kernel.org>]

On Tue, Nov 03, 2026 at 08:12:48PM +0200, Edward Srouji wrote:
> mlx5 devices recognise RDMA traffic on receive with a fixed classifier: UDP
> destination port 4791. Packets that match are handed to the RDMA receive path
> (NIC_RX_RDMA); everything else continues down the normal NIC RX path.
> 
> RDMA transports that identify their packet differently don't reach the
> RDMA receive path as they use a different packet format entirely.
> 
> This series adds NIC_RX_CLASS, a new per-vport receive steering domain placed
> ahead of NIC_RX/NIC_RX_RDMA. It gives a privileged RDMA transport manager a
> programmable classification stage that supports matching on arbitrary header
> fields and marking the packet as RDMA.
> 
> NIC_RX_CLASS reuses the model used by RDMA_TRANSPORT_RX/TX domains:
> - one root namespace per vport, owned by the eswitch manager;
> - targeted from userspace by `ib_port`, as with the existing per-vport
>   domains;
> - gated by `rdma_transport_manager`, and by
>   `rdma_transport_manager_other_eswitch` for the cross-eswitch (LAG /
>   socket-direct) case;
> - restricted to callers holding the same UCAP that per-port matchers already
>   require, so an unprivileged process cannot program another vport's
>   classification.
> 
> A rule in this domain takes no destination: it matches, attaches a
> modify-header action to mark the packet as RDMA and explicitly allows it
> onward. ALLOW is the only terminal action the domain accepts; a flow counter
> may be attached alongside it.
> 
> Patch 1 adds the IFC bits, patch 2 fixes ALLOW handling in the FTE merge
> logic, patch 3 unifies the per-vport root namespace init/cleanup helpers,
> and patch 4 adds the NIC_RX_CLASS domain itself. Patch 5 exposes the new
> domain and the ALLOW action to userspace.
> 
> Patches 1-4 are mlx5_core changes and are targeted at the shared mlx5-next
> branch, prefixed accordingly; patch 5 depends on them and is targeted at
> rdma-next.
> 
> Signed-off-by: Edward Srouji <edwards@nvidia.com>
> ---
> Michael Gur (5):
>       net/mlx5: Add NIC_RX_CLASS IFC bits
>       net/mlx5: fs, treat ALLOW as a conflicting action upon merge
>       net/mlx5: fs, refactor RDMA transport manager root namespace init/cleanup
>       net/mlx5: fs, add NIC_RX_CLASS steering domain support
>       RDMA/mlx5: Expose NIC_RX_CLASS flow table type to userspace

I applied this to mlx5-next and rdma-next.

Thanks