Current Limitations
Check support boundaries before choosing a deployment role.
Check support boundaries before choosing a deployment role.
rustbgpd is a public alpha. It is suitable for lab, data-center fabric pilots, IX route-server pilots, and programmable control-plane deployments where the operator is comfortable with an evolving config and gRPC API.
The exception is the deliberately narrow, machine-inventoried v1 route-server / route-reflector / controller contract. That promise applies only to the listed IPv4/IPv6 unicast RS/RR, scoped RR-only, and FlowSpec controller surfaces; it does not promote the rest of the project out of alpha.
Routing role
- Linux FIB integration is opt-in and scoped. RFC 7999 BLACKHOLE discard routes
and configured
[[fib_tables]]unicast route installation are available, with per-peer / peer-group allow-lists and per-table route-count caps. ECMP, per-class ECMP caps,multipath_relax, and Link Bandwidth weighted multipath are opt-in. Full router parity still needs broader redistribution policy and non-BGP route-manager scope. - VPNv4 / VPNv6 support is route-reflector / controller-feed only. VRF import, MPLS label forwarding, and CE-facing attachment are out of scope.
- BGP-LS support is receive, reflection, API export, and ORR topology input. rustbgpd does not synthesize local BGP-LS objects. ORR computes only the RFC 9552 default topology: non-default MT-ID and malformed topology inputs are excluded fail-closed, while Flex-Algorithm data is ignored and no selectable non-default/Flex SPF is implemented. The BGP-LS trust boundary is per-scope, not per-peer: ORR unions BGP-LS Adj-RIB-In across all peers into one default topology, so any negotiated BGP-LS speaker can inject well-formed default-topology nodes or links that shift another ORR client's best-path selection. This is inherent to the RFC 9107 union model — the MT/Flex isolation above scopes topology inputs, not speakers.
- FlowSpec (SAFI 133, IPv4 and IPv6) is control plane only: rules are
received, injected, and distributed, but rustbgpd installs no FlowSpec
filters in a dataplane. The scoped controller contract
covers local origination, reconciliation, and committed advertised views,
and the guarded mutations
of ADR-0144 fence a controller's
late-committing adds and deletes with mutation windows and rule revisions.
Both pin the API shape only, with no forwarding guarantee. RFC 9117 feasibility validation of received rules remains alpha and
is opt-in (
[flowspec] validation = "rfc9117"; the default is"off"); see ADR-0135. IPv4/IPv6 VPN FlowSpec (SAFI 134) supports opt-in controller origination with explicit RD/RT scope, outside v1; see the VPN API contract. Received VPN rules are decoded and discarded, with no default-unicast feasibility lookup or RT-Constrain filtering. rustbgpd creates no local VRF, and local acceptance does not prove router installation or traffic enforcement. Nonzero IPv6 prefix offsets use RFC 8956 pattern framing and do not interoperate with peers that use the older full-prefix framing, including GoBGP 4.11.0 and earlier rustbgpd releases. Depending on the length and offset, a nonzero-offset rule from such a peer either fails to decode and resets the whole session with NOTIFICATION 3/9 every time it is sent, or decodes cleanly with a different, bit-shifted match (for example length 32, offset 1). An established session does not prove compatibility. Keep IPv6 FlowSpec rules with a nonzero prefix offset (global or VPN) withdrawn on both sides until every peer uses RFC 8956 framing; offset-zero IPv6 and IPv4 rules are unaffected. The qualification receipt separates that comparator failure from the passing core checks. - Confederations are not implemented.
- Conditional advertisement (ADR-0137)
is alpha and outside the v1 inventory. It covers IPv4 and IPv6 unicast only,
[[dynamic_neighbors]]ranges do not accept attachments, and an attached neighbor leaves update-group sharing. See Conditional advertisements. - RFC 5004 (prefer the existing external best path) is not implemented. Path
selection is deterministic: below the eBGP-over-iBGP step, ties are broken
by the lowest effective BGP Identifier, then the shorter CLUSTER_LIST, then
the lowest peer address, so a newly learned external path that ties the
current best down to the identifier and carries the lower identifier
replaces it. BIRD and OpenBGPD default to the same behavior. If an operator
asks, the change would take the shape of an opt-in preference for the
existing external path; see the RFC 4271 §9.1.2.2 notes in
docs/reference/rfc-notes.md. - Transparent route-server export preserves the accepted route's next hop.
Operators can opt into fail-closed pre-policy ownership validation with
next_hop_ownership = "strict_peer", which accepts unicast announcements only when the complete wire next-hop identity is the advertising session's own address. Same-AS alternate next hops and explicit authorization remain deferred; see ADR-0107. SLAT-driven translation for mixed RFC 8950 client fleets is not implemented; ADR-0128 keeps it demand-gated. Implementation requires either working-group adoption or named IXP demand, plus a real SLAT producer and retained fixtures. A uniform IPv6 fleet needs no translation:rs-config-renderrenders an arouteserverrfc8950IPv6 session as a dual-family session withstrict_peer, and refuses the shape as soon as an IPv4-session member is present.
EVPN
EVPN Types 1–5 route reflection is a
scoped RR-only v1 role;
the rest of EVPN remains alpha. Its local VTEP and dataplane support are
Linux/VXLAN-only.
The reflector implements scoped
SRv6 service framing
and eligibility checks,
with unchanged-next-hop reflection of eligible raw attributes. SRv6 PE import,
service origination, next-hop rewriting, and forwarding remain unimplemented.
Inspection views show an optional
reconstructed_sid
when a single route supplies an unambiguous Function transposition;
cross-route Argument composition is not reconstructed, and the value does not
affect selection.
See SRv6 route troubleshooting
for the retained-but-ineligible route case.
Shipped and interop-tested:
- Route-reflector mode for EVPN route types 1-5.
- Bidirectional single-homed L2VNI VTEP: remote-MAC FDB programming, local MAC-only / MAC+IP / SVI Type 2 origination, Type 3 IMET, and push-notified sub-second convergence.
- Symmetric Interface-less IRB / Type 5 with transactional L3 FIB programming.
- IPv6 VXLAN underlay: IPv6 VTEP addresses, IPv6 BGP transport, 16-octet EVPN
next hops, and remote-VTEP FDB rows with an IPv6
dst. - Active-active multi-homing building blocks: DF election, Type 1/4, BUM
suppression, aliasing ECMP via FDB nexthop groups, and all-active Type 5
receive. Preference-based DF election can opt into RFC 9785 Don't-Preempt
(
df_dont_preempt); the M123 receipt qualifies Highest-Preference Don't-Preempt restart, recovery and reference-withdrawal behavior against SR Linux 25.10.1 on one two-PE all-active segment, without forwarding.esi = "auto-lacp"derives a Type 1 ESI from a Linux bond's LACP partner. A daemon started with no[[ethernet_segments]]takes its first segment, explicit-ESI orauto-lacp, live through SIGHUP orApplyEvpnRuntime. Type 2 (STP) ESI derivation is not implemented. See Ethernet Segments. - Duplicate-MAC detection with quarantine, bounded key-only status listing, and manual clear. The listing intentionally exposes no detector clocks or recovery durations and caps each atomic snapshot response at 4096 rows.
- Controller injection for EVPN route types 2, 3, and 5.
- VNI-per-broadcast-domain VLAN-aware bridge support, including SVD / collect-metadata VXLAN detection and programming.
- Opt-in managed bridge, fixed-VNI VXLAN, SVD / collect-metadata VXLAN, VLAN
upper, VRF, and fixed-VNI L3VXLAN lifecycle under
[managed_netdevs]. - Non-zero Ethernet Tag VLAN-aware-bundle receive/reflect in RR mode.
- Single-homed VLAN-aware bundle members on the VTEP: per-member Ethernet Tag Type 2 and Type 3 origination and per-tag import onto the member's bridge VLAN, with global VNIs only.
Type 6 SMET typed receive, reflection, withdrawal, and inspection are also implemented in the alpha RR lane. The M113 controlled raw-peer proof checks reflection, withdrawal, and error recovery with an independent TShark decoder. This does not establish vendor interoperability or scale; the earlier receipts above cover other route types. See the SMET boundary.
Known EVPN gaps:
- L3VNI/device/table IP-VRF identity changes remain restart-required by design. Supported decomposable EVPN runtime edits commit live in ordered primitive steps; unsupported dependency cycles fail closed before commit, and residual mid-sequence convergence failures fail-stop on the last committed generation.
- VLAN-aware bundle VTEP members have a bounded
SR Linux 25.10.1 import and forwarding receipt.
Single-homed members can bind an
ip_vrfonly ininterface_lessmode, using ESI0, Ethernet Tag0Type 5 routes with a separate L3VNI. The M120 local receipt qualifies this shape against SR Linux 25.10.1. Tag-scoped multi-homing, locally assigned VNIs, non-zero Ethernet Tag Type 5, and GW-IP/ESI overlay-index recursion through bundle members remain unsupported. The binding does not provision Linux SVI/VRF attachments or add Type 2 symmetric host routing. EVPN VTEP and IRB behavior remains alpha. - VXLAN local-bias split-horizon for all-active shared segments remains an ASIC/offload-dependent limitation of the Linux softswitch path. RFC 9746 §2.2 makes local bias the only split-horizon mechanism for VXLAN (the ESI Label Split Horizon Type MUST be 00 for tunnel type 8), so there is no ESI-label alternative to implement for this lane.
- All-active CE forwarding requires an Ethernet Segment
interfacebinding. On all-active segments, eligible peer Type 2 routes install owned local-port FDB rows, including on a non-DF PE that never learned the CE MAC. This does not originate peer-only MAC/MAC-IP routes or program peer-only IP adjacencies. Unbound segments retain control-plane behavior without this forwarding guarantee. Single-active segments retain ordinary local MAC relearning and do not install these peer-synchronized rows. Later CE source traffic on the same port does not promote an owned peer-sync row to local origination; Linux retains the external-learning marker. Concurrent foreign FDB writers can also race the pre-apply ownership check; see the ADR-0085 amendment. - A shared all-active ES with stock FRR 10.7.1 on Linux is qualified only
with
highest-preferenceand FRR as the DF. FRR becomes non-DF on every VNI on algorithm disagreement, and it does not filter BUM when it is non-DF: its netlink provider never applies the non-DF bridge-port update. With rustbgpd as DF beside it, the CE receives each BUM frame twice, once through each PE. The mixed-peer receipt covers one rustbgpd PE, one FRR PE and one FRR remote VTEP in a single all-active ES; it is not scale evidence. See FRR all-active ES peers. - Symmetric Interface-less IRB does not implement
RTA_VIA, so a Type 5 prefix and its next hop must be the same address family. Under an IPv6 VTEP only IPv6 tenant prefixes are carried: an IPv4 prefix is refused at origination and dropped at import rather than installed against a cross-family gateway. The same constraint applies with the families reversed. An IPv6 VXLAN underlay is otherwise supported for the L2 path and for same-family symmetric IRB. The receipts covering it are single-homed, so a multi-homed IPv6 underlay is untested rather than known unsupported. - EVPN route types 7–11 are not decoded or reflected. Unsupported typed NLRIs are discarded on receive under RFC 7606 §5.4, including on a reflector; they never enter the RIB. The per-type discard counter and connection-scoped warnings identify the affected peer and type.
- SMET origination, IGMP/MLD proxy procedures, and multicast forwarding are not implemented. PBB-EVPN, multicast EVPN service, MPLS/SRv6 encapsulation, VPWS, and E-Tree are demand-shaped rather than part of the current VXLAN/Linux lane.
See docs/project/evpn-enablement.md and the EVPN ADRs for the
full gate ladder.
Transport and session features
- TCP-AO supports ordered static-neighbor and direct dynamic-prefix keyrings on Linux. SIGHUP can append a non-preferred successor, later select it and observation-gate predecessor deprecation, and later delete deprecated MKTs that are neither Current nor RNext. Key edits/reordering, selected or non-deprecated-key deletion, and protected-owner CRUD require a daemon restart.
- TCP MD5 and GTSM are supported.
- BFD supports IPv4, IPv6 global, and interface-scoped IPv6 link-local
single-hop asynchronous sessions, and IPv4 / IPv6 global multihop (RFC 5883)
asynchronous sessions, for static neighbors. Sessions accepted through a
[[dynamic_neighbors]]range run IPv4 and IPv6 global single-hop non-strict sessions configured on the range; strict and multihop range BFD are not supported. Multihop sessions apply no receive TTL bound (RFC 5883 leaves a GTSM-style check optional; none is configurable) and cannot target an IPv6 link-local neighbor. Echo, demand mode and authentication remain follow-up work. BGP sessions to non-adjacent peers require no separate enablement and can be distance-bounded withttl_security_hops. - BGP unnumbered supports static IPv6 link-local neighbors, including IPv4 unicast over RFC 8950, and Linux interface-only neighbors discovered through Router Advertisements or inbound BGP connections. See interface-only neighbors for discovery prerequisites and shared-address restrictions.
- Link-Local Next Hop capability 77 (
draft-ietf-idr-linklocal-capability-06) is experimental and off by default. It is enabled per neighbor or peer group withlink_local_next_hop; there is no global default. Only explicitly interface-bound IPv6 link-local neighbors may enable it, and[[dynamic_neighbors]]ranges cannot. With or without the capability, a link-local next hop is exported only on the interface where it was learned; a route that would carry it elsewhere is withheld and counted inbgp_exact_export_rejections_total{reason="missing_ipv6_next_hop"}unless next-hop self orlocal_ipv6_nexthoprewrites it.
Configuration control
- The peer-manager private command lane and config-bridge replacement lane are
lossless capacity-one FIFO channels. A planning, apply, rollback, or reload
operation already using one of these lanes serializes later control work;
callers wait instead of commands being dropped or an unbounded queue growing.
Under a slow configuration operation this can increase concurrent management
request latency even though routing-session work continues on its own actor
paths. Live operator reads such as neighbor snapshots and RIB queries use a
separate bounded read lane and are admitted at defined points during policy
and reload generations (ADR-0132), but
they keep their deadlines and can still time out behind an individual
synchronous work unit. See
configuration reload for the
admission points.
GetPolicyStats(rbgp policy stats) instead reads counter rosters published by the peer manager and RIB manager (ADR-0136) and does not queue on either actor; it can still wait for a pending session publication or a busy counter or dataset error lock under its one absolute two-second deadline, and returnsDEADLINE_EXCEEDEDwith no partial rows when that deadline expires. A fleet import read can opt intoallow_partial, which skips closed import-session publications and lists them inincomplete_peer_addressesinstead of failing; seeGetPolicyStats.
Operational proof
Published benchmarks and receipts cover the main protocol, interop, scale, and memory surfaces. Multi-day soak automation beyond the archived harnesses remains future work.
See: