Re: [PATCH 2/3] NETFILTER module xt_hmark, new target for HASH based fwmark

Hi Hans,

On Mon, Jan 23, 2012 at 10:49:16AM +0100, Hans Schillstrom wrote:

On Monday 23 January 2012 10:12:41 Pablo Neira Ayuso wrote:

quoted

On Mon, Jan 23, 2012 at 12:20:15AM +0100, Hans Schillstrom wrote:

quoted

The text should clarify that this is valid for the fragments not the "flow"

quoted

I've got one scenario that may break with this assumption:

1) your traffic follows one path over router A and B to reach your
   firewall F which requires no fragmentation at all.

I missed the last part here  "requires no fragmentation at all"

quoted

quoted

quoted

2) path to router B becomes broken while there are established flows
   with firewall F.
3) router A decides to forward packets to router C, which fragment
   packets because it is using smaller MTU than router A.
4) packets arrive to firewall F, then hashing is calculated based on
   addresses, not ports, and you load-sharing becomes inconsistent.

This can rarely happen, but it does, it would break.

To fix this, I think that HMARK requires that you have to specify the
hashing strategy. If you want to support fragments, use only
addresses. If you're sure you will not get fragments, use layer 3 and
layer 4 information.

This can be acomplished by setting --hmark-sp-mask and --hmark-dp-mask to Zero
Then you don't use port in the hash calc.

OK, it would be great if we can provide a cleaner interface. The
current behaviour uses layer3-layer4 tuple hashing plus defaulting to
layer3 in case of fragments.

I'd prefer explicit configuration options:

--hashmark-method layer3
        use only address for hashing, this is fragment safe.

--hashmark-method layer3-layer4
        use addresses and ports for hashing, fragments not supported
        unless defrag is enabled.

Still, if you want to support the current behaviour, it should be
something like:

--hashmark-method layer3-layer4-fragments
        use addresses and ports for hashing, for fragments default to
        layer3 hashing. Document scenario in which hash consistency
        may break.

The behaviour of the target has to be specified by the configurations.
Defaulting to internal assumptions seems obscure to me.