On Tue, 23 Jul 2019 19:12:23 +1000
"Trent W. Buck" <[email protected]> wrote:
> (I hope these questions aren't too stupid for this list!)
>
> A stateful firewall ruleset usually starts like:
>
> ct state established,related accept
> ct state invalid drop
> # Hooray, 99% of packets are now decided;
> # we only need to think hard the ones in ct state new!
Personally, I drop INVALID packets in PREROUTING in order to waste no more CPU
cycles than are necessary because we already know that that packet cannot be
processed.
>
> If you want to guard against a coworker "accidentally" adding NOTRACK in
> raw, you might tweak that to drop "ct untracked" as well:
>
> ct state established,related accept
> ct state != new drop
>
> But nftables has these "verdict map" things now:
>
> ct state vmap { established:accept, related:accept, invalid:drop }
>
> Is that "better" than the original?
> It's one less rule, so it ought to be more efficient, right?
>
> One obvious downside is you lose the ability to counter/log/reject the
> ct state invalid packets, because those aren't valid in verdict_expr.
>
>
> Why is the original allowed to have "established,related" without braces?
> [UPDATE: answered below]
>
> I'm a little bit worried because "nft describe ct state" looks like a
> list of flags (powers of two), and in parser_bison.y I don't understand
> ct_expr, but set_flag_list is using a comma as a bitwise OR.
>
> Can a packet be *both* established and related at the same time?
'RELATED' is the first ('NEW') packet of a conn that has been determined to be
related to an existing conn. As far as I know, once a conn transitions from
'RELATED' to 'ESTABLISHED', the system loses all knowledge that the conn was
ever 'RELATED' in the first place. I mention this because I've never been able
to determine if netfilter will close/terminate 'RELATED' conns when it
closes/terminates the master. For example, if netfilter terminates an FTP
control conn, does it also terminate data conns that were 'RELATED' to that
control conn?
>
> If so, is "ct state established,related" matching any packet that has
> established or related *OR BOTH*?
>
> If so, does that mean it will accept some packets that my vmap (above) won't?
>
> The parser won't accept "{related,established: accept}" in a vmap, but
> it will accept a pipe. Does this has the same meaning as the original
> pair of rules?
>
> ct state vmap { invalid : drop, established | related : accept }
>
> I think so, because the pipe comes back out as a comma here:
>
> # nft add rule 'inet x y ct state established|related accept'
> # nft list chain inet x y
> table inet x
> chain y {
> ct state established,related accept
> }
> }
>
> Oh using the English word "or" is probably clearer, at least for
> anglophones:
>
> ct state established or related accept
> ct state vmap { established or related: accept, invalid: drop }
>
> The equivalence of "or" and "|" and "," --- except inside a set/map ---
> wasn't obvious to me from the nft manpage (as at 0.9.1). I "caught on"
> because equivalences like "ne" and "!=" were mentioned here:
> https://wiki.nftables.org/wiki-nftables/index.php/Building_rules_through_expressions
> (no bitwise operators there, though)
>
>
> PS: AFAICT if a vmap doesn't match, the rule doesn't match, so
> processing just carries on to the next rule. Is there a way to have a
> "default value" in the vmap? e.g.
>
> ct state vmap { established or related: accept,
> new: continue,
> OTHERWISE: return }
>
> I can't see it in verdict_map_list_expr or in the manpage.
>
> PPS: my earlier "chain comment" question, I now realize the "NOOP rule"
> I wanted is just "continue", e.g.
>
> table inet x {
> chain y {
> continue comment "This chain is the common start for INPUT and
> FORWARD."
> continue comment "It allows the things you (almost always) want."
> ct state vmap { established or related: accept; invalid: drop }
> iiftype loopback accept
> icmp type { ... } accept
> icmpv6 type { ... } accept
> }
> }
>
