docs: log crowdsec engine restart-policy fix and 2.5wk firewall-bouncer auth outage (2026-09-03)

running_log.md

@@ -25,3 +25,49 @@
 - Confirmed via Uptime Kuma dashboard: mail.infernalaquatics.com back to "Up".
 
 **Takeaway for next time:** When touching NPM's forward_host in the DB directly (vs. through the UI), always re-save through the UI (or explicitly restart+verify the generated `.conf` file) to guarantee the config regenerates — a raw container restart is not sufficient to pick up DB-only changes.
+
+## 2026-09-03 — CrowdSec down after reboot + long-standing firewall-bouncer auth failure (Net-Controller)
+
+**Symptom:** CrowdSec showing "Down" in Uptime Kuma after the earlier full host reboot (same reboot that broke mailcow reachability, see previous entry).
+
+### Issue 1 — CrowdSec engine container not running
+
+**Investigation:**
+- `docker ps -a` showed `crowdsec` container `Exited (0)` ~17 hours prior — a clean shutdown (`SIGTERM received, shutting down`), not a crash.
+- `docker inspect crowdsec` confirmed exit code `0`, no OOM kill, no error — it simply received the reboot's shutdown signal and never came back.
+- Root cause: `docker inspect --format '{{.HostConfig.RestartPolicy.Name}}'` showed `RestartPolicy: no` — unlike the rest of the stack, this container had no auto-restart policy configured, so it stayed down after the reboot instead of self-healing like every other service on the host.
+- Container's compose label pointed to `/data/compose/14`, which no longer exists on disk (stale label from original provisioning, likely Coolify/Portainer-managed at some point) — not pursued further since fixing the running container directly was simpler and safer.
+
+**Fix:**
+- `docker start crowdsec`
+- `docker update --restart unless-stopped crowdsec` — ensures it now survives future reboots automatically.
+- Verified: container came up clean, re-downloaded/updated its hub (parsers, scenarios, collections — flagged one outdated collection `crowdsecurity/http-cve`, auto-updated to CVE-2022-44877 v0.4), started processing normally, confirmed pulling from CrowdSec Central API.
+
+### Issue 2 — firewall-bouncer crash-looping since ~Aug 18 (pre-existing, unrelated to reboot)
+
+**Discovery:** While verifying CrowdSec's engine logs, noticed the `crowdsec-firewall-bouncer` client was receiving `403 Forbidden` on every single `/v1/decisions/stream` poll (every 10s), both before and after the reboot — meaning this was a standing issue, not caused by tonight's outage.
+
+**Investigation:**
+- The bouncer is **not** a docker container — it's a host-level systemd service: `crowdsec-firewall-bouncer.service` (installed via `crowdsec-firewall-bouncer-iptables` package, config at `/etc/crowdsec/bouncers/crowdsec-firewall-bouncer.yaml`).
+- `systemctl status` showed it `activating (auto-restart)` — crash-looping continuously.
+- `journalctl -u crowdsec-firewall-bouncer.service` showed the same fatal error on every cycle: `level=fatal msg="process terminated with error: bouncer stream halted"`, immediately followed by systemd restarting it (10s backoff).
+- **Scale of the problem:** restart counter was at **5,829** by the time this was caught, with log rotation files dating back to **Aug 18, 2026** — meaning this bouncer had been non-functional (crashing every ~10s) for roughly **2.5 weeks** before discovery.
+- Root cause confirmed via `docker exec crowdsec cscli bouncers list` on the engine side — **zero bouncers registered**. The API key configured in the bouncer's yaml (`api_key: MRPnN0XqxrVOspREia3c4jvRYfi+/Cr/9bUKAcxdoWg`) did not correspond to any bouncer known to the CrowdSec engine (likely invalidated by a prior CrowdSec engine reset/reinstall, or the bouncer was never properly registered after an update).
+- **Security impact:** with the bouncer unable to authenticate, ban decisions from CrowdSec's engine (local detections + CrowdSec Central API community blocklist) were **never being enforced at the firewall (nftables) layer** for the ~2.5 week window. CrowdSec's detection/logging continued working the whole time (it runs independently), but actual IP blocking was not happening.
+
+**Fix:**
+1. Generated a fresh bouncer API key on the engine: `docker exec crowdsec cscli bouncers add firewall-bouncer-netcontroller` → new key `DkIEnS1TfGfly49gAAb4cbn1T6/+EZBVSEFCoEojt0M`.
+2. Backed up the existing config: `/etc/crowdsec/bouncers/crowdsec-firewall-bouncer.yaml.bak.<timestamp>`.
+3. Replaced the `api_key:` line in `/etc/crowdsec/bouncers/crowdsec-firewall-bouncer.yaml` with the new key via `sed`.
+4. `sudo systemctl restart crowdsec-firewall-bouncer.service`.
+
+**Verification:**
+- `systemctl is-active` → `active` and stable (previously dying in under a second on every cycle; now running continuously with no restarts).
+- `cscli bouncers list` on the engine now shows `firewall-bouncer-netcontroller` with `Valid: ✔️`, IP `172.18.0.1`, auth type `api-key`, and a fresh `Last API pull` timestamp.
+- CrowdSec engine logs confirmed the bouncer's `/v1/decisions/stream` polls now return `HTTP 200` (was `403`) every 10s as expected.
+- Confirmed actual enforcement is live: `sudo nft list ruleset` shows the `ip crowdsec` and `ip6 crowdsec6` tables (matching the `nftables:` config block — note: NOT under the `inet` family, took one wrong guess before finding the right table family) with active, incrementing packet/byte counters on the block chains (17,990+ packets processed at time of check).
+
+**Takeaways for next time:**
+- CrowdSec's own container needed an explicit `RestartPolicy` fix (`unless-stopped`) that most of the other stack containers already had — worth spot-checking other standalone/non-compose-managed containers on this host for the same gap.
+- The firewall-bouncer is a **host-level systemd service**, not a docker container — easy to overlook when doing container-focused health checks (`docker ps` alone won't show it). Add it to the regular health-check rotation going forward.
+- Bouncer API keys are one-time-display (`cscli bouncers add` only shows the key once) and are independent of the CrowdSec engine's own restarts — if the engine's bouncer registry is ever wiped/reset, all bouncer services depending on old keys will silently start failing until manually re-registered. Consider a monitoring check on `cscli bouncers list` validity or on the bouncer's own systemd restart count as an early-warning signal for this specific failure mode.