Fix rekey dual-initiation, parent selection, peer reconnect, and control socket

Rekey dual-initiation race (FMP + FSP):
When both sides' rekey timers fire simultaneously on high-latency links
(Tor ~700ms RTT), both msg1s cross in flight before dampening can
suppress the second initiation. Each node acts as both initiator and
responder, with set_pending_session() from the responder path destroying
the initiator's in-progress state. Each side ends up with a
pending_new_session from a different Noise handshake, causing AEAD
verification failure after cutover and link death. Fix: deterministic
tie-breaker in both FMP msg1 handler and FSP SessionSetup handler
(smaller NodeAddr wins as initiator). Also guard against pending session
overwrite from retransmitted msg1s.

Parent selection SRTT gate:
Peers without MMP RTT data are excluded from parent candidacy via
has_srtt() filter, preventing freshly reconnected peers from appearing
artificially attractive. First-RTT triggers immediate parent re-eval.
The gate in evaluate_parent() skips unmeasured candidates when any peer
has cost data, but falls back to default cost 1.0 during cold start
when no peer has MMP data yet.

Auto-reconnect on all peer removal paths:
The excessive-decryption-failure and peer-restart epoch-mismatch removal
paths were missing schedule_reconnect() calls, causing auto-connect peers
to be permanently abandoned after those events.

Control socket accessibility:
Socket and parent directory chowned to root:fips with mode 0770/0750 so
group members can use fipsctl/fipstop without root.

Log level changes:
Default RUST_LOG changed to debug in systemd service files. "Unknown
session index" log reduced from debug to trace.
This commit is contained in:
Johnathan Corgan
2026-03-16 00:51:55 +00:00
parent 324535e76d
commit 5f7fe989f3
12 changed files with 214 additions and 17 deletions
+14 -3
View File
@@ -340,7 +340,15 @@ impl TreeState {
if coords.contains(&self.my_node_addr) {
continue;
}
let cost = peer_costs.get(peer_id).copied().unwrap_or(1.0);
// If any peer has MMP cost data, only consider measured peers.
// This prevents freshly connected peers (no SRTT, default cost 1.0)
// from appearing artificially cheap. During cold start (no peer has
// MMP data, peer_costs is empty), fall back to default cost 1.0.
let cost = match peer_costs.get(peer_id) {
Some(&c) => c,
None if peer_costs.is_empty() => 1.0,
None => continue,
};
let eff_depth = coords.depth() as f64 + cost;
match &best_peer {
None => best_peer = Some((*peer_id, eff_depth)),
@@ -396,11 +404,14 @@ impl TreeState {
// --- Same root, cost-aware comparison with hysteresis ---
// Current parent's effective_depth
// Current parent's effective_depth.
// If peer_costs is non-empty but current parent has no entry,
// treat as maximally expensive so any measured candidate can win.
// If peer_costs is empty (cold start), use default cost 1.0.
let current_parent_cost = peer_costs
.get(self.my_declaration.parent_id())
.copied()
.unwrap_or(1.0);
.unwrap_or(if peer_costs.is_empty() { 1.0 } else { f64::INFINITY });
let current_parent_coords = self.peer_ancestry.get(self.my_declaration.parent_id());
let current_parent_eff = match current_parent_coords {
Some(coords) => coords.depth() as f64 + current_parent_cost,