Re: [PATCH nf-next 0/3] netfilter: nf_tables: reject loads from uninitialized registers

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Sun, May 07, 2023 at 01:22:54PM +0200, Florian Westphal wrote:
> Pablo Neira Ayuso <pablo@netfilter.org> wrote:
> > On Fri, May 05, 2023 at 04:51:13PM +0200, Florian Westphal wrote:
> > > Pablo Neira Ayuso <pablo@netfilter.org> wrote:
> > > > On Fri, May 05, 2023 at 01:16:53PM +0200, Florian Westphal wrote:
> > > > > Keep a per-rule bitmask that tracks registers that have seen a store,
> > > > > then reject loads when the accessed registers haven't been flagged.
> > > > > 
> > > > > This changes uabi contract, because we previously allowed this.
> > > > > Neither nftables nor iptables-nft create such rules.
> > > > > 
> > > > > In case there is breakage, we could insert an 'store 0 to x'
> > > > > immediate expression into the ruleset automatically, but this
> > > > > isn't done here.
> > > > > 
> > > > > Let me know if you think the "refuse" approach is too risky.
> > > > 
> > > > Might the NFT_BREAK case defeat this approach? Sequence is:
> > > > 
> > > > 1) expression that writes on register hits NFT_BREAK (nothing is written)
> > > > 2) expression that read from register, it reads uninitialized data.
> > > >
> > > > From ruleset load step, we cannot know if the write fails, because it
> > > > is subject to NFT_BREAK.
> > > 
> > > Yes, but its irrelevant: If 1) issues NFT_BREAK, 2) won't execute.
> > 
> > And register tracking is done per rule, given context is per rule too,
> > good.
> > 
> > I wonder if it is worth to move the bitmask away from nft_ctx, given
> > this structure is stored in the struct nft_trans, hence increasing the
> > size of this object which is not required at a later state, maybe
> > there is a need for a new container structure that store data useful
> > for the initial preparation step of the commit protocol.
> 
> Hmm, this will get messy.
> 
> I only see two alternatives:
> 
> - place the bitmask in the pernet structure.
> - add struct nft_expr_ctx as a container structure, which has
>   nft_ctx as first member and the bitmask as second member, to
>   be used for NEWRULE and NEWSETELEM instead of nft_ctx.

Can the 'level' field be moved to this nft_expr_ctx structure? This
field is only used from the preparation phase (not in the commit
phase).

Probably we need to rename nft_ctx to nft_trans_ctx, so it contains
the fields that are needed from the commit phase. Then, re-add a
nft_ctx again which contains nft_trans_ctx at the beginning, then the
register bitmap and the level field. Thus, any future fields only
required by preparation phase only will go in nft_ctx, and fields that
are specifically are set up from preparation phase and consumed from
commit step go in nft_trans_ctx.

It is a bit of churn, but it is probably good to tidy up this for
future extensions?

Let me know, thanks.