Firewall

Last updated

UFW + CrowdSec, with a strict division of labor.

Division of labor

ConcernOwner
Which ports are open at allUFW
Which IPs may reach an open portCrowdSec decisions
Which IPs are temporarily blocked because of bad behaviorCrowdSec scenarios + decisions
WAF (request-content inspection)CrowdSec AppSec
Per-user outbound egressnftables + cgroup v2 vmap

UFW is demoted: no ufw deny from <ip> rules. The CrowdSec bouncers consult decisions in real time and reject at the IP layer (nftables for the network bouncer, app-level for nginx/Stalwart bouncers).

Port baseline

Standard rules the installer applies:

ALLOW 22/tcp     # SSH / SFTP
ALLOW 25/tcp     # SMTP MTA
ALLOW 53/tcp     # PDNS
ALLOW 53/udp     # PDNS
ALLOW 80/tcp     # HTTP (incl. ACME HTTP-01)
ALLOW 443/tcp    # HTTPS
ALLOW 465/tcp    # SMTP submission TLS
ALLOW 587/tcp    # SMTP submission STARTTLS
ALLOW 993/tcp    # IMAPS
ALLOW 995/tcp    # POP3S (only if POP3 is enabled in Stalwart)
DENY  IN ALL     # default deny
ALLOW OUT ALL    # outbound left open by default (constrained per-user by nftables)

Add / remove ports at /jabali-admin/security → UFW.

Blocklists

Besides its own detections, CrowdSec pulls the free CrowdSec community blocklist (CAPI) and the bouncers drop the IPs on it. The admin can add more lists (see Extra blocklists). Security → CrowdSec → Blocklists shows how many active decisions each source holds.

The panel does not use the CrowdSec console (app.crowdsec.net): console enrolment was removed, and jabali update turns console alert forwarding off on hosts that had enrolled. Blocklists that are subscribed through the console are not received.

IPs imported with cscli decisions import are enforced like any other decision and show under Blocklists as cscli-import. The panel does not remove them; they expire with the duration given at import.

Extra blocklists

The Extra blocklists card on the same tab adds public lists on top of the community blocklist. All are off by default.

  • FireHOL level 1: Spamhaus DROP hijacked networks, DShield’s top attacking networks and Feodo botnet controllers. Very few false positives.
  • FireHOL level 2: addresses seen attacking in about the last 48 hours (blocklist.de, DShield, GreenSnow). Some false positives.
  • FireHOL level 3: addresses reported in the last 30 days (BruteForceBlocker, CI Army, DShield, VXVault and others). Catches more, with more false positives.
  • IPsum level 3: addresses found on at least 3 of the 30 or more public attack and malware lists IPsum merges each day. Few false positives.
  • Spamhaus DROP (IPv6): IPv6 networks hijacked or run by spammers and cybercriminals. The IPv4 DROP list is part of FireHOL level 1. Very few false positives.
  • Your own lists: up to 10 https URLs of plain-text lists, one IP or CIDR per line (# and ; comments are skipped). The URL may not point at a private address or at this server.

The panel downloads each list that is on once a day and bans every entry, whole networks included, on every port (websites, mail and SSH) for 48 hours from that download, so a host whose panel is down for a day keeps its lists. These are never banned: private and reserved ranges, entries wider than /8 (IPv4) or /29 (IPv6), this server’s own addresses, CrowdSec allowlist entries, and the addresses Google and Bing publish for their crawlers. A listed network that holds one of these addresses is left out whole. If the crawler addresses cannot be loaded and no earlier copy is on the server, the lists are not refreshed that day and their bans stay until they expire.

Turning a list off, or removing one of your own, lifts its bans when the run finishes, within a minute or two. An address added to a CrowdSec allowlist is let in within a few seconds, and banned again within a few seconds of its removal if a list still holds it; an entry that expires is banned again on the next refresh. A visitor wrongly listed stays blocked until you turn the list off or allowlist them. The card shows, for each list, how many entries the last run banned and how many it left out and why.

These bans are not CrowdSec decisions, so they do not show under Blocklists. The agent loads them into its own nftables table, which drops traffic from the listed addresses on input and forward:

nft list table inet jabali_lists

The agent rebuilds the table when it starts, so the lists survive a reboot. If the CrowdSec allowlists cannot be read, the table is left as it was, and the card says the firewall was not updated.

Earlier versions imported the lists as CrowdSec decisions (origin cscli-import, scenario jabali-list:<list>), which banned only the first address of each network. After the update, those of a list that is still on expire within 48 hours; those of a list that is off are deleted.

Migrating from old ufw deny rules

If you migrated from a setup that used UFW for IP banning:

jabali ufw migrate-ip-bans

This walks ufw status numbered, lifts every from <ip> deny rule into a CrowdSec decision with reason migrated-from-ufw, then deletes the UFW rule. Idempotent, safe to re-run.

Per-user egress

See security.md. The default policy is “outbound 443 + mail submission to localhost only, everything else dropped”; admin overrides per-user.

What about IPv6?

Same port rules apply (ufw default deny incoming, manual allow per port). CrowdSec scenarios match v4 and v6. AppSec is protocol-agnostic.