Re: [nf-next PATCH v2] netfilter: nf_tables: Introduce NFTA_RULE_ACTUAL_EXPR

[Phil Sutter <phil@nwl.cc>]

On Sat, Feb 04, 2023 at 10:41:37AM +0100, Pablo Neira Ayuso wrote:
> On Fri, Feb 03, 2023 at 05:21:29PM +0100, Phil Sutter wrote:
> [...]
> > On Fri, Feb 03, 2023 at 04:32:01PM +0100, Pablo Neira Ayuso wrote:
> [...]
> > > I also wonder if this might cause problems with nftables and implicit
> > > sets, they are bound to one single lookup expression that, when gone,
> > > the set is released. Now you will have two expressions pointing to an
> > > implicit set. Same thing with implicit chains. This might get tricky
> > > with the transaction interface.
> > 
> > While indeed two lookup expressions will refer to the same anonymous
> > set, only one of those expressions will ever be in use. There's no way
> > the kernel would switch between rule variants (or use both at the same
> > time).
> 
> OK, but control plane will reject two lookup expressions that refer to
> the same anonymous set.

Only if it sees the second expression: If NFTA_RULE_ACTUAL_EXPR is
present, the kernel will copy the content of NFTA_RULE_EXPRESSIONS into
a buffer pointed to by nft_rule::dump_expr. It does not inspect the
content apart from nla_policy checking which merely ensures it's a
nested array of elements conforming to nft_expr_policy (i.e., have a
NAME and DATA attribute).

The copied data is touched only by nf_tables_fill_rule_info() which
copies it as-is into the skb. Later, nf_tables_rule_destroy() just frees
the whole blob.

So effectively the kernel doesn't know or care what expressions are
contained in NFTA_RULE_EXPRESSIONS.

> > > iptables is rather simple representation (no sets), but nftables is
> > > more expressive.
> > 
> > That's not true, at least ebtables' among match is implemented using
> > sets. :)
> 
> Then better have a look at this implicit set scenario I describe above
> because I cannot see how this can work.

Sure, I'll give it a try.

Cheers, Phil