[lance0]

Summary

The example configs/rustbgpd.toml, which the Compose rustbgpd profile mounts, ships a placeholder neighbor (192.0.2.1) that never connects, and it left Graceful Restart at rustbgpd's default (on). Injected routes are held from selection after a speaker restart until its Graceful Restart peers send End-of-RIB or the restart window (gr_restart_time, 120 seconds by default) expires. A placeholder that never comes up holds them for the full window, so every quickstart or lab restart left re-injected FlowSpec rules at selected: false for about two minutes. MarkReplayComplete does not end this hold.

• configs/rustbgpd.toml: set graceful_restart = false on the placeholder neighbor and keep the neighbor so the file still works as a template. A comment says to enable GR on real peers that support it.
• docs/reference/speaker-backends.md, docs/how-to/upgrading.md: note the post-restart selection hold, that MarkReplayComplete does not end it, and why the shipped example disables GR. The wording matches the selection-hold text that main gained with the selection_deferred export-evidence reason, and the speaker-backends note links to that section.
• CHANGELOG.md: [Unreleased] / ### Fixed entry.

The key is the neighbor-level graceful_restart (bool, default true) in rustbgpd v0.80.1. It appears in rustbgpd's configuration reference, which has a "To disable GR for a specific peer" example, and in the v0.80.1 rustbgpd --dump-config-schema output (Neighbor.graceful_restart). The hold is bounded by gr_restart_time. gr_eor_grace_time is a separate setting: it holds End-of-RIB toward a peer, is off by default and is unchanged here.

Left unchanged on purpose

• lab/interop/configs/rustbgpd.toml and rustbgpd-srlinux-probe.toml peer with real FRR / SR Linux nodes that come up, and the FRR lab tests GR behaviour directly (test_graceful_restart_stale_path).
• tests/common/rustbgpd.rs: its never-Established unicast peer is part of the scenario under test, and fixture restarts use stop_with_timeout(Some(0)) (SIGKILL). That restart writes no planned-restart marker, so no deferral starts.
• docker-compose.yml and scripts/test-compose.py have no neighbor config of their own. Compose mounts the fixed example.

Verification

• rustbgpd --check --strict with the pinned v0.80.1 image digest from tests/fixtures/rustbgpd/releases.json: config OK (exit 0) for both the old and the new configs/rustbgpd.toml.
• Restart timing, using a disposable rustbgpd v0.80.1 container (pinned digest) with the shipped config, a bind-mounted runtime dir and the same cpuset. Steps: inject an IPv4 FlowSpec rule over the UDS, then docker restart -t 30 (SIGTERM, a planned restart that writes the GR marker). Then re-inject and poll the local injection view (ListFlowSpecRoutes with received_peer_address = "0.0.0.0") until the rule is selected: true:
  • Old config (negative control): re-injected at +0.27 s, selected: true at +119.9 s. The Loc-RIB was empty until then. The daemon logged RFC 4724 route-selection deferral released ... reason: Selection_Deferral_Timer expired for IPv4 and IPv6 FlowSpec, after a marker with restart_time_secs: 120.
  • New config: re-injected at +0.25 s, selected: true at +0.32 s. The rule was in the Loc-RIB right after re-injection.
  • The test containers and runtime dirs were removed afterwards.
• just gate on db4685a: passed (exit 0).
• git diff --check: clean.