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

[Florian Westphal]

Linux regression tracking (Thorsten Leemhuis) <regressions@xxxxxxxxxxxxx> wrote:
> On 12.09.23 00:57, Pablo Neira Ayuso wrote:
> > 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.
> > [...]
> 
> Hmmmm. Well, this sounds like a kernel regression to me that normally
> should be dealt with on the kernel level, as users after updating the
> kernel should never have to update any userspace stuff to continue what
> they have been doing before the kernel update.

This is a combo of a userspace bug and this new sanity check that
rejects the incorrect ordering (adding rules to the already-bound
anonymous chain).

nf_tables uses a transaction allor-nothing model, this means that any
error that occurs during a transaction has to be reverse/undo all the
pending changes.  This has caused a myriad of bugs already.

So while this can be theoretically fixed in the kernel I don't see
a sane way to do it.  Error unwinding / recovery from deeply nested
errors is already too complex for my taste.

> Can't the kernel somehow detect the incorrect bytecode and do the right
> thing(tm) somehow?

Theoretically yes, but I don't feel competent enough to do it, just look
at all the UaF bugs of the past month.