Articles

Test the firewall before you trust it

jailops firewall test says whether a new connection between two jails, or from an address or the host, would pass the firewall, and which rule decides at each step. How to read it, how to use it in a script before a change, what it cannot know, and how to check the answer with a real connection.

Why ask first

A refused connection tells you nothing.

The usual test of a firewall rule is to try the connection and see. When it fails you know that something said no, not which rule, which policy or which of the two anchors a connection between two jails passes. Then you read the policies, the project's mandatory policy, the managed policy of an app and the defaults, and work it out.

jailops firewall test does that reading for you, on the rules JailOps would load, and prints the rule that decided. It changes nothing and sends no packet.

The console has the same test as a picture, with a form for the two ends: See a connection before it happens.

Asking

From, to, and a port.

The example: a project policy default (deny in, ping allowed) for every jail, and a policy postgres on the database jail that allows the address set postgresql-clients to port 5432. The set holds jail:web.

three questions
$ jailops firewall test --from jail:web --to jail:db --port 5432
$ jailops firewall test --from jail:web --to jail:db --port 5433
$ jailops firewall test --from jail:other --to jail:db --port 5432

An end is a jail of the project (jail:web), one address, or host. --proto is tcp by default; udp needs a port too, and icmp and icmp6 take --icmp-type (a ping, an echo request, when you leave it out).

Reading the answer

One line per anchor the connection meets.

The first question prints this:

web to db, port 5432
tcp port 5432 10.20.0.11 -> 10.20.0.10 (inet): PASS
  web sends: allowed by the default of the anchor
  db takes: allowed by policy postgres rule 1 "who may reach PostgreSQL"

The first line is the verdict. Under it is one line for each end of the connection: what the source sends and what the destination takes, and the rule that decided. Here the web jail may send anything, so its own anchor allows it by its default, and the database anchor allows it because of rule 1 of policy postgres, with the comment you wrote for it. That comment is why a reviewer can read this without opening the policy.

Add -v when you want the detail: the pf line each answer was taken from, the rules that were skipped and what the test does not cover.

The second question, port 5433, ends in db takes: blocked by the default of the anchor: no rule allows it, so the default of the jail, deny in, decides. The third is blocked the same way, because jail:other is not in the set. The names that can decide are a policy and its rule number, a project policy, the managed policy of an app, the isolation between projects, and the default of the anchor.

The exit status is 0 for pass, 1 for block and 8 when the test cannot say. A usage error is 2.

The two ends

A connection between two jails has two hops.

What a jail sends is judged on its own port. What reaches the destination leaves the host on the gateway interface of its network. So a connection between two jails has to pass the source's anchor and the destination's, and the test shows both. From the host or from an address outside, only the destination side is evaluated.

  • Another project is an external host: its jails are addresses here, and only the other end's anchor is evaluated. Write the address (--from 10.30.0.5), as you do in the policy.
  • An address that belongs to a jail of the project is that jail.
  • A set is not an end. Name a jail or an address and the set will show up as the rule that let it in.
  • IPv4 and IPv6 are tested each on their own, for every family the two ends have in common (--family 4 or 6 to pick one).

In a script

Write down what must hold, and check it after a change.

The exit status makes it a check. Keep a short file of the connections that must pass and the ones that must not, and run it after every change of a policy or a set:

firewall-check.sh
#!/bin/sh
# 0 = pass, 1 = block; anything else is an error or "cannot say" and fails the check
fail=0
expect() {  # expect pass|block FROM TO PORT
  want=$1; shift
  jailops firewall test --from "$1" --to "$2" --port "$3" >/dev/null 2>&1; rc=$?
  case "$want:$rc" in pass:0|block:1) ;; *) echo "FAIL: $want $1 -> $2:$3 (exit $rc)"; fail=1;; esac
}
expect pass  jail:web   jail:db 5432
expect block jail:web   jail:db 5433
expect block jail:other jail:db 5432
exit $fail

-o json prints the same answer for a program: the verdict, each case and hop with the policy, the rule number and its comment, and the caveats. A console that simulates a change can use the same answer later.

What it cannot know

It says so, and does not guess.

  • A connection that is already open. pf keeps state, so the answer is for a new connection.
  • A raw PF fragment in front of the decision. The answer is unknown (exit status 8), naming the policy that has it.
  • The host's own pf.conf and NAT. The test evaluates the JailOps anchors; what the host's rules do before or after them is stated as not covered.
  • ICMP types. Without --icmp-type the test is an echo request, and the answer says so.

"Pass" therefore means "as far as these rules go", and -v lists under Not covered what that leaves out.

Checking with a real connection

The same question, asked to the network.

When it matters, ask the network too. From the host, with nc if the jail has it:

a real connection
$ jailops exec web -- nc -z -w 2 10.20.0.10 5432; echo $?

The answer should agree with the test. If the two ever disagree, the test is wrong, and that is a bug worth reporting.

The rules, sets and policies the test reads are explained in A firewall you can read, and why a project is a boundary in Projects are boundaries.

More articles

Try it on one server.

Try the console from the screenshots, install on a clean FreeBSD 15.1 machine, or add JailOps to a host you already run.