Bgp crashes

[donaldsharp]

see individual commits, but stop 2 shutdown crashes that happening in the topotests

[greptile-apps]

Greptile Summary

This PR fixes two shutdown-order crashes in the BGP EVPN multi-homing code. First, it adds bgp_evpn_es_route_table_purge() to drain leftover bgp_path_info entries from each ES route table before bgp_table_unlock() is reached in bgp_evpn_es_free(), preventing a crash when the pi_hash is non-empty at teardown. Second, it adds a NULL guard on bgp_vrf->l2vnis in bgpevpn_unlink_from_l3vni() to handle the race where VRF EVPN cleanup has already called list_delete(&bgp->l2vnis) (which NULLs the pointer) before late VNI teardown fires.

Confidence Score: 4/5

Safe to merge; both changes are tightly scoped shutdown-path fixes with no impact on the data-plane code path.

Both fixes follow established patterns already present in the codebase (same reap loop, same list_delete NULL-after-free convention) and address confirmed topotest crashes. No new logic is introduced in normal (non-shutdown) operation. Score is 4 rather than 5 because the purge function accesses pi->peer inside bgp_path_info_reap during a phase where the BGP instance may be partially torn down, but this same risk exists in the surrounding callers and the check if (pi->peer) is already present.

bgpd/bgp_evpn_mh.c — confirm bgp_path_info_reap is safe when called after bgp instance teardown (peer pointers may be stale).

Important Files Changed

Filename	Overview
bgpd/bgp_evpn_mh.c	Adds bgp_evpn_es_route_table_purge() to drain path-info entries from the ES route table during shutdown, called from bgp_evpn_mh_finish() before bgp_table_unlock(); logic is a safe duplicate of the existing bgp_evpn_es_route_del_all() inner loop.
bgpd/bgp_evpn_private.h	Guards listnode_delete() with a NULL check on bgp_vrf->l2vnis to prevent crash when the list has already been freed (and set to NULL by list_delete()) during VRF teardown before late VNI cleanup runs.

Sequence Diagram

sequenceDiagram
    participant MH as bgp_evpn_mh_finish
    participant Purge as bgp_evpn_es_route_table_purge (NEW)
    participant Reap as bgp_path_info_reap
    participant ESFree as bgp_evpn_es_free
    participant Table as bgp_table_unlock

    MH->>Purge: for each ES in rb_tree
    Purge->>Purge: loop over dest/pi in route_table
    Purge->>Reap: bgp_path_info_mark_for_delete + reap
    Reap-->>Purge: dest (assert non-NULL)
    MH->>MH: force-cleanup ES-EVIs
    MH->>ESFree: bgp_evpn_es_local_info_clear → es_free
    ESFree->>Table: bgp_table_unlock (pi_hash now empty ✓)

    note over MH,Table: Without the purge, pi_hash non-empty crash here

Loading

_{Reviews (1): Last reviewed commit: "bgpd: Prevent crash when list is already..." | Re-trigger Greptile}

[ton31337]

LGTM