Session

Segment Routing over IPv6 (SRv6, RFC 8986) lets an application or operator encode a packet-processing program directly in the IPv6 header. SRv6 has been supported in the Linux kernel since release 4.10, and a rich open-source ecosystem (FRR, SONiC, Cilium, VPP) has grown on top of it. Following the SRv6 workshops at Netdev 0x16 (Lisbon 2022) and 0x19 (Zagreb 2025), this edition covers three directions that directly impact the kernel datapath — source-routed AI backends, L2 services beyond VXLAN, and provider-grade SRv6 deployment with service protection — plus a short ecosystem update.

Source-routed AI backend networks - AI backends are tightly controlled fabrics where ECMP-based spreading gives little direct control over how flows map onto the fabric. Source routing with SRv6 enables deterministic, congestion-aware path placement decided at the host/NIC, removing hash collisions and reacting to link/switch failures without waiting for control-plane convergence. Our group has been exploring this direction [1], which also underpins the recently announced industry MRC protocol (covered in a main-track talk). Here we focus on what this means for Linux: efficient per-packet segment-list selection in seg6/seg6local, the interplay of eBPF/XDP and NIC offload at line rate, and the supporting role of FRR, SONiC and telemetry.

SRv6 L2 services beyond VXLAN - RFC 8986 defines the L2 endpoint behaviors (End.DX2/DT2U/DT2M) and RFC 9252 the EVPN overlay over SRv6. Kernel L2 support is still limited to the End.DX2 cross-connect, with no native L2 endpoint netdevice. A recent netdev RFC series adds End.DT2U and the sr6 Ethernet pseudowire device, enabling a VXLAN-like deployment model. We discuss the path to multipoint services, EVPN-over-SRv6 in FRR, and replacing VXLAN in cloud orchestrators (Kubernetes CNIs, OVN, OpenStack Neutron).

SRv6 for network providers and service protection - Beyond the data-center and emerging use cases above, SRv6 is gaining traction in telco/provider networks. This contribution presents an end-to-end provider-style design built entirely on the GNU/Linux and FRRouting stack: addressing and per-service VRF design, SRv6 SID allocation, BGP and IS-IS configuration, and traffic engineering — with the same end-to-end services also realized in SR-MPLS on the same network to contrast the coexistence, configuration and operation of the two technologies. To troubleshoot such networks it introduces tablesnoop [2], a lookup-level observability tool for live tracing of policy-based routing, IPv4/IPv6 route lookups, SRv6 head-end and endpoint behaviors, and (SR-)MPLS label operations (swap/push/pop), filling the gap between coarse header-level capture (tcpdump) and verbose kernel function-call tracing (pwru, retis, ipftrace2). Finally, it extends Linux SRv6 programming with Redundancy Protection (including DetNet use cases): a new R-SID format (draft-ietf-spring-sr-redundancy-protection) and SR policy head-end behaviors, with the IEEE 802.1CB FRER implementation XDPFRER [3] extended for SRv6 encapsulation. The session shows how XDP and the existing SRv6 routing stack are combined, and measures the performance impact relative to unprotected SRv6 forwarding.

We close with a short update on the SRv6 open-source ecosystem: the latest kernel features and refactorings, FRR enhancements and SONiC support.

[1] C. Filsfils, P. Camarillo, A. Abdelsalam, A. Quinci, A. Tulumello, A. Mayer, P. Loreti, L. Bracciale, S. Salsano, “Toward Deterministic Path Placement in AI Backends: A Practical SRv6-Based Architecture”, IEEE CNSM 2025.

[2] tablesnoop, https://github.com/EricssonResearch/tablesnoop

[3] XDPFRER, https://github.com/EricssonResearch/xdpfrer