Re: [PATCH net] ipv6: Prevent ipv6_find_hdr() from returning ENOENT for valid non-first fragments

From: Rahul Sharma
Date: Mon Jan 12 2015 - 23:23:51 EST


Hi

On Mon, Jan 12, 2015 at 5:21 PM, Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx> wrote:
> On Mon, Jan 12, 2015 at 04:38:16PM +0530, Rahul Sharma wrote:
>> Hi Pablo, Hannes
>>
>> On Fri, Jan 9, 2015 at 9:20 PM, Hannes Frederic Sowa
>> <hannes@xxxxxxxxxxxxxxxxxxx> wrote:
>> > On Fr, 2015-01-09 at 12:45 +0100, Pablo Neira Ayuso wrote:
>> >> Hi Hannes,
>> >>
>> >> On Fri, Jan 09, 2015 at 12:34:15PM +0100, Hannes Frederic Sowa wrote:
>> >> > On Fri, Jan 9, 2015, at 08:18, Rahul Sharma wrote:
>> >> > > Hi Pablo,
>> >> > >
>> >> > > On Fri, Jan 9, 2015 at 5:35 AM, Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx>
>> >> > > wrote:
>> >> > > > On Thu, Jan 08, 2015 at 11:39:16PM +0100, Hannes Frederic Sowa wrote:
>> >> > > >> Hi Pablo,
>> >> > > >>
>> >> > > >> On Thu, Jan 8, 2015, at 21:53, Pablo Neira Ayuso wrote:
>> >> > > >> > I'm afraid we cannot just get rid of that !ipv6_ext_hdr() check. The
>> >> > > >> > ipv6_find_hdr() function is designed to return the transport protocol.
>> >> > > >> > After the proposed change, it will return extension header numbers.
>> >> > > >> > This will break existing ip6tables rulesets since the `-p' option
>> >> > > >> > relies on this function to match the transport protocol.
>> >> > > >> >
>> >> > > >> > Note that the AH header is skipped (see code a bit below this
>> >> > > >> > problematic fragmentation handling) so the follow up header after the
>> >> > > >> > AH header is returned as the transport header.
>> >> > > >> >
>> >> > > >> > We can probably return the AH protocol number for non-1st fragments.
>> >> > > >> > However, that would be something new to ip6tables since nobody has
>> >> > > >> > ever seen packet matching `-p ah' rules. Thus, we restore control to
>> >> > > >> > the user to allow this, but we would accept all kind of fragmented AH
>> >> > > >> > traffic through the firewall since we cannot know what transport
>> >> > > >> > protocol contains from non-1st fragments (unless I'm missing anything,
>> >> > > >> > I need to have a closer look at this again tomorrow with fresher
>> >> > > >> > mind).
>> >> > > >>
>> >> > > >> The code in question is guarded by (_frag_off != 0), so we are
>> >> > > >> definitely processing a non-1st fragment currently. The -p match would
>> >> > > >> happen at the time when the packet is reassembled and thus ipv6_find_hdr
>> >> > > >> will find the real transport (final) header at this point (I hope I
>> >> > > >> followed the code correctly here).
>> >> > > >
>> >> > > > Then, Rahul should get things working by modprobing nf_defrag_ipv6.
>> >> > >
>> >> > > I already had nf_defrag_ipv6 installed when the issue occured. But I
>> >> > > see ip6table_raw_hook returning NF_DROP for the second fragment.
>> >> >
>> >> > That's what I expected. I think the change only affects hooks before
>> >> > reassembly.
>> >>
>> >> reassembly happens at NF_IP6_PRI_CONNTRACK_DEFRAG (-400), so that
>> >> happens before NF_IP6_PRI_RAW (-300) in IPv6 which is where the raw
>> >> table is placed.
>> >
>> > I tried to reproduce it, but couldn't get non-1st fragments getting
>> > dropped during traversal of the raw table. They get dropped earlier at
>> > during reassembly or pass.
>> >
>> > I agree with Pablo, I also would like to see more data.
>> >
>> > Thanks,
>> > Hannes
>> >
>> >
>>
>> I enabled pr_debug() and there was no error in nf_ct_frag6_gather().
>> It seems to have defragmented the packet correctly. As expected,
>> ipv6_defrag() returns NF_STOLEN for the first packet after queuing it.
>> For the next fragment, ipv6_defrag() calls nf_ct_frag6_output() after
>> after reassembling it.
>
> nf_ct_frag6_output() doesn't exist anymore. You're using an old
> kernel, you should have started by telling so in your report.
>
> See 6aafeef ("netfilter: push reasm skb through instead of original
> frag skbs").

I apologize for not mentioning the kernel version in my first mail. I
had suspected problem in ipv6_find_hdr, the code for which was same.
Anyway, thanks for the help. I ll try to figure out how to make this
work in my kernel.

Thanks,
Rahul
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/