On Sat, Jul 25, 2020 at 10:31:01PM -0400, Gunnar Niels wrote:

I've been continuing to experiment with firewalld and learn the underlying userspace tools like iptables and nft for expressing firewall rules, but from what I've been reading, it seems like nftables is intended to replace iptables, which is possibly heading towards deprecation (if anyone can shed some light on that, would be much appreciated)?

Yes. That's accurate.

Maybe this blog will help: https://firewalld.org/2018/07/nftables-backend

If that's true, it appears that firewalld whever it is in use still needs to work on top of *both* iptables and nftables. When I add a direct rule, that gets added directly to the cooresponding iptables chain, while rich rules look like they are added to nftable chains.

Also correct. With the nftables backend all the firewalld primitives: services, ports, forward-ports, masquerade, etc use nftables. But direct rules will still use iptables. The --direct feature is really a thin wrapper around iptables/ebtables. firewalld will _not_ translate iptables into nftables. That's a task doomed for failure as it would never be 100% correct.

I'm working on my Arch Linux workstation, but I'm also responsible for a number RHEL/CentOS boxes that I'd like to upgrade to their respective v8 editions, but I'd really like a rock solid understanding of firewalld and the systems underneath before I'm comfortable in productions.

So how do iptables and nftables work in conjunction with one another when they seem like they potentially could hold conflicting rules? Is there some order of precedence where these are ultiately boiled down to a single golden source of rules within the kernel?

I think the "Behavioral Changes" section of the above blog post covers this. It's also covered in the CAVEATS section of man page firewalld.direct(5).

Below is copy/pasted from the man page.

-->8--

CAVEATS

Depending on the value of FirewallBackend (see firewalld.conf(5)) direct rules behave differently in some scenarios.

Packet accept/drop precedence Due to implementation details of netfilter inside the kernel, if FirewallBackend=nftables is used direct rules that ACCEPT packets don't actually cause the packets to be immediately accepted by the system. Those packets are still be subject to firewalld's nftables ruleset. This basically means there are two independent firewalls and packets must be accepted by both (iptables and nftables). As an aside, this scenario also occurs inside of nftables (again due to netfilter) if there are multiple chains attached to the same hook - it's not as simple as iptables vs nftables.

There are a handful of options to workaround the ACCEPT issue:

1. Rich Rules

If a rich rule can be used, then they should always be preferred over direct rules. Rich Rules will be converted to the enabled FirewallBackend. See firewalld.richlanguage(5).

2. Blanket Accept

Users can add an explicit accept to the nftables ruleset. This can be done by adding the interface or source to the trusted zone.

This strategy is often employed by things that perform their own filtering such as: libvirt, podman, docker.

Warning: This means firewalld will do no filtering on these packets. It must all be done via direct rules or out-of-band iptables rules.

3. Selective Accept

Alternatively, enable only the relevant service, port, address, or otherwise in the appropriate zone.

4. Revert to the iptables backend

A last resort is to revert to the iptables backend by setting FirewallBackend=iptables. Users should be aware that firewalld development focuses on the nftables backend.

For direct rules that DROP packets the packets are immediately dropped regardless of the value of FirewallBackend. As such, there is no special consideration needed.

Firewalld guarantees the above ACCEPT/DROP behavior by registering nftables hooks with a lower precedence than iptables hooks.

Direct interface precedence With FirewallBackend=iptables firewalld's top-level internal rules apply before direct rules are executed. This includes rules to accept existing connections. In the past this has surprised users. As an example, if a user adds a direct rule to drop traffic on destination port 22 existing SSH sessions would continue to function, but new connections would be denied.

With FirewallBackend=nftables direct rules were deliberately given a higher precedence than all other firewalld rules. This includes rules to accept existing connections.