Re: Regression: Commit "netfilter: nf_tables: disallow rule addition to bound chain via NFTA_RULE_CHAIN_ID" breaks ruleset loading in linux-stable

[Salvatore Bonaccorso <carnil@debian.org>]

Hi Timo,

On Tue, Sep 12, 2023 at 01:39:59PM +0200, Timo Sigurdsson wrote:
> Hi Pablo,
> 
> Pablo Neira Ayuso schrieb am 12.09.2023 00:57 (GMT +02:00):
> 
> > Hi Timo,
> > 
> > On Mon, Sep 11, 2023 at 11:37:50PM +0200, Timo Sigurdsson wrote:
> >> Hi,
> >> 
> >> recently, Debian updated their stable kernel from 6.1.38 to 6.1.52
> >> which broke nftables ruleset loading on one of my machines with lots
> >> of "Operation not supported" errors. I've reported this to the
> >> Debian project (see link below) and Salvatore Bonaccorso and I
> >> identified "netfilter: nf_tables: disallow rule addition to bound
> >> chain via NFTA_RULE_CHAIN_ID" (0ebc1064e487) as the offending commit
> >> that introduced the regression. Salvatore also found that this issue
> >> affects the 5.10 stable tree as well (observed in 5.10.191), but he
> >> cannot reproduce it on 6.4.13 and 6.5.2.
> >> 
> >> The issue only occurs with some rulesets. While I can't trigger it
> >> with simple/minimal rulesets that I use on some machines, it does
> >> occur with a more complex ruleset that has been in use for months
> >> (if not years, for large parts of it). I'm attaching a somewhat
> >> stripped down version of the ruleset from the machine I originally
> >> observed this issue on. It's still not a small or simple ruleset,
> >> but I'll try to reduce it further when I have more time.
> >> 
> >> The error messages shown when trying to load the ruleset don't seem
> >> to be helpful. Just two simple examples: Just to give two simple
> >> examples from the log when nftables fails to start:
> >> /etc/nftables.conf:99:4-44: Error: Could not process rule: Operation not
> >> supported
> >>                         tcp option maxseg size 1-500 counter drop
> >>                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> >> /etc/nftables.conf:308:4-27: Error: Could not process rule: Operation not
> >> supported
> >>                         tcp dport sip-tls accept
> >>                         ^^^^^^^^^^^^^^^^^^^^^^^^
> > 
> > I can reproduce this issue with 5.10.191 and 6.1.52 and nftables v1.0.6,
> > this is not reproducible with v1.0.7 and v1.0.8.
> > 
> >> Since the issue only affects some stable trees, Salvatore thought it
> >> might be an incomplete backport that causes this.
> >> 
> >> If you need further information, please let me know.
> > 
> > Userspace nftables v1.0.6 generates incorrect bytecode that hits a new
> > kernel check that rejects adding rules to bound chains. The incorrect
> > bytecode adds the chain binding, attach it to the rule and it adds the
> > rules to the chain binding. I have cherry-picked these three patches
> > for nftables v1.0.6 userspace and your ruleset restores fine.
> 
> hmm, that doesn't explain why Salvatore didn't observe this with
> more recent kernels.
> 
> Salvatore, did you use newer userspace components when you tested
> your 6.4.13 and 6.5.2 builds?

It does explain now because understanding the issue better. While one
while experinting should only change each one constraint for the
6.4.13 and 6.5.2 testing I indeed switched to a Debian unstable
system, which has newer userpace nftables and so not triggering the
issue. This was missleading for the report.

> As for the regression and how it be dealt with: Personally, I don't
> really care whether the regression is solved in the kernel or
> userspace. If everybody agrees that this is the best or only viable
> option and Debian decides to push a nftables update to fix this,
> that works for me. But I do feel the burden to justify this should
> be high. A kernel change that leaves users without a working packet
> filter after upgrading their machines is serious, if you ask me. And
> since it affects several stable/longterm trees, I would assume this
> will hit other stable (non-rolling) distributions as well, since
> they will also use older userspace components (unless this is
> behavior specific to nftables 1.0.6 but not older versions). They
> probably should get a heads up then.

So if it is generally believed on kernel side there should not happen
any further changes to work with older userland, I guess in Debian we
will need to patch nftables. I'm CC'ing Arturo Borrero Gonzalez
<arturo@debian.org>, maintainer for the package. The update should go
ideally in the next point releases from October (and maybe released
earlier as well trough the stable-updates mechanism).

FWIW: In Debian bullseye we have 0.9.8 based nftables, in bookworm
1.0.6, so both will need those fixes.

As 0ebc1064e487 is to address CVE-2023-4147 other distros picking the
fix will likely encounter the problem at some point. It looks Red Hat
has taken it (some RHSA's were released), I assume Ubuntu will shortly
as well release USN's containing a fix.

Regards,
Salvatore