Re: [PATCH nft] tests: shell: check that rule add with index works with echo

From: Pablo Neira Ayuso <pablo@netfilter.org>
To: Phil Sutter <phil@nwl.cc>,
	netfilter-devel@vger.kernel.org, Eric Garver <eric@garver.life>
Subject: Re: [PATCH nft] tests: shell: check that rule add with index works with echo
Date: Wed, 4 Sep 2019 21:31:07 +0200	[thread overview]
Message-ID: <20190904193107.pwghs7oqhdzanqeu@salvia> (raw)
In-Reply-To: <20190904191718.kzgbqdsgdjctqqli@salvia>

On Wed, Sep 04, 2019 at 09:17:18PM +0200, Pablo Neira Ayuso wrote:
> On Wed, Sep 04, 2019 at 10:13:37AM +0200, Phil Sutter wrote:
> > Pablo,
> > 
> > On Tue, Sep 03, 2019 at 07:27:13PM -0400, Eric Garver wrote:
> > > If --echo is used the rule cache will not be populated. This causes
> > > rules added using the "index" keyword to be simply appended to the
> > > chain. The bug was introduced in commit 3ab02db5f836 ("cache: add
> > > NFT_CACHE_UPDATE and NFT_CACHE_FLUSHED flags").
> > > 
> > > Signed-off-by: Eric Garver <eric@garver.life>
> > > ---
> > > I think the issue is in cache_evaluate(). It sets the flags to
> > > NFT_CACHE_FULL and then bails early, but I'm not sure of the best way to
> > > fix it. So I'll start by submitting a test case. :)
> > 
> > In 3ab02db5f836a ("cache: add NFT_CACHE_UPDATE and NFT_CACHE_FLUSHED
> > flags"), you introduced NFT_CACHE_UPDATE to control whether
> > rule_evaluate() should call rule_cache_update(), probably assuming the
> > latter function merely changes cache depending on current command. In
> > fact, this function also links rules if needed (see call to
> > link_rules()).
> > 
> > The old code you replaced also did not always call rule_cache_update(),
> > but that was merely for sanity: If cache doesn't contain rules, there is
> > no point in updating it with added/replaced/removed rules. The implicit
> > logic is if we saw a rule command with 'index' reference, cache would be
> > completed up to rule level (because of the necessary index to handle
> > translation).
> > 
> > I'm not sure why you introduced NFT_CACHE_UPDATE in the first place, but
> > following my logic (and it seems to serve no other purpose) I would set
> > that flag whenever NFT_CACHE_RULE_BIT gets set. So IMHO,
> > NFT_CACHE_UPDATE is redundant.
> 
> Please, just go ahead simplify this in case you found a way to do it.

IIRC, the idea was:

* NFT_CACHE_UPDATE tells to update the cache incrementally.
* NFT_CACHE_RULE tells to fetch the rule cache.

I think you're right, they currently overlap, because if
NFT_CACHE_RULE is requested, then NFT_CACHE_UPDATE necessarily needs
to happen, right?

Oh, there's one scenario where this is not the case: If flush is
requested, then the NFT_CACHE_RULE flag is set off, while the
NFT_CACHE_UPDATE is still left in place.