Re: [PATCH] netfilter: NFT_SOCKET don't use NF_SOCKET_IPV6 without NF_TABLES_IPV6

On Tue, Jul 10, 2018 at 11:10:40AM +0200, Arnd Bergmann wrote:

On Tue, Jul 10, 2018 at 10:05 AM, Máté Eckl [off-list ref] wrote:

quoted

On Tue, Jul 10, 2018 at 10:02:27AM +0200, Máté Eckl wrote:

quoted

On Mon, Jul 09, 2018 at 11:35:09PM +0200, Arnd Bergmann wrote:

quoted

It is now possible to build the nft_socket module as built-in when
NF_TABLES_IPV6 is disabled, and have NF_SOCKET_IPV6=m set manually.

In this case, the NF_SOCKET_IPV6 functionality will be useless according
to the explanation in commit 35bf1ccecaaa ("netfilter: Kconfig: Change
IPv6 select dependencies"), but on top of that it also causes a link
error:

net/netfilter/nft_socket.o: In function `nft_socket_eval':
nft_socket.c:(.text+0x162): undefined reference to `nf_sk_lookup_slow_v6'

This changes the compile-time check so we don't attempt to use
the NF_SOCKET_IPV6 code when it cannot be used, and make it all
compile again. That may lead to unexpected behavior when a user
enables NF_SOCKET_IPV6 but cannot use it, but seems to be the
logical conclusion of the 35bf1ccecaaa change.

Fixes: 35bf1ccecaaa ("netfilter: Kconfig: Change IPv6 select dependencies")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>

I think this should be fixed in the Kconfig rather than inside the module(s).

Should we revert your patch then, or do you have a better idea?

quoted

quoted

I did some investigation and it turns out that you missed a circumstance. This
link error occures only if NFT_SOCKET=y && NF_SOCKET_IPV6=m && NF_TABLES_IPV6=y
(cannot be m here if NFT_SOCKET is y).

No, if NF_TABLES_IPV6=y the problem cannot happen, since NFT_SOCKET then
selects NF_SOCKET_IPV6=y as well. Before your patch, it would always select
NF_SOCKET_IPV6 when it could, so it worked in all configurations.

Sorry I wanted to write NF_TABLES_IPV6=n...
So: NFT_SOCKET=y && NF_SOCKET_IPV6=m && NF_TABLES_IPV6=n causes linkage error.
NFT_SOCKET=m && NF_SOCKET_IPV6=m && NF_TABLES_IPV6=n compiles fine.

quoted

quoted

 And probably the same with
iptables-related modules. Probably this possibility should be eliminated.

NF_TPROXY_IPV6 might be in the same situation.

I tried coming up with a combination that is broken for NF_TPROXY_IPV6=m
but could not. From what I can see with

config NETFILTER_XT_TARGET_TPROXY
       tristate '"TPROXY" target transparent proxying support'
       depends on IP6_NF_IPTABLES || IP6_NF_IPTABLES=n
       select NF_TPROXY_IPV6 if IP6_NF_IPTABLES

and

#if IS_ENABLED(CONFIG_IP6_NF_IPTABLES)

inside of net/netfilter/xt_TPROXY.c, there is no way we can end up with
xt_TPROXY calling into the nf_tproxy_ipv6 loadable module from
a built-in context. This is the same approach I used in my patch,
just with IP6_NF_IPTABLES instead of NF_SOCKET_IPV6, in both
the Kconfig dependency and the module.

Right, I see your point. I have an alternative solution which seems more robust
to me, but I might be overthinking this situation.

So
Accepted-by: Máté Eckl [off-list ref]

      Arnd