Message-ID: <53CB8848.7050400@redhat.com>
Date:	Sun, 20 Jul 2014 11:13:44 +0200
From:	Daniel Borkmann <dborkman@...hat.com>
To:	Vlad Yasevich <vyasevich@...il.com>
CC:	davem@...emloft.net, jgunthorpe@...idianresearch.com,
	netdev@...r.kernel.org, linux-sctp@...r.kernel.org
Subject: Re: [PATCH net] net: sctp: inherit auth_capable on INIT collisions

On 07/19/2014 04:23 AM, Vlad Yasevich wrote:
> On 07/18/2014 07:03 PM, Daniel Borkmann wrote:
>> On 07/19/2014 12:13 AM, Daniel Borkmann wrote:
>>> On 07/18/2014 11:59 PM, Vlad Yasevich wrote:
>>>> On 07/18/2014 03:17 PM, Daniel Borkmann wrote:
>>>>> On 07/18/2014 04:38 PM, Vlad Yasevich wrote:
>>>>> ...
>>>>>> Why is the original value of asoc->peer.auth_capable = 0?
>>>>>> In case of collision, asoc is the old association that
>>>>>> existed on the system.  That association was created as part of
>>>>>> sending the INIT.  If it is processing a duplicate COOKIE-ECHO
>>>>>> as you say, then it has already processed the INIT-ACK and
>>>>>> should have determined that the peer is auth capable.
>>>>>>
>>>>>> Thus the capability of the new and the old associations should
>>>>>> be same if we are in fact processing case B (collision).
>>
>> What I can see is the following that leads to this situation:
>>
>> 1) asoc A sends the INIT, goes from CLOSED into COOKIE_WAIT
>> 2) asoc B receives it, calls into sctp_sf_do_5_1B_init() where it
>>     actually creates asoc B, responds with INIT_ACK, goes from CLOSED
>>     into COOKIE_WAIT
>
> I think this is a race.  asoc B doesn't exist yet.  we have a listening
> socket that responds normally to the INIT-ACK.  The next thing that happens
> is the app initiates a connection thus creating asoc B and triggering INIT.
>
>> 3) asoc A receives INIT, thus collision, calls into sctp_sf_do_5_2_1_siminit()
>> 3.1) asoc A calls into sctp_sf_do_unexpected_init(), creates a temp asoc,
>>       does sctp_process_init() on the temp asoc (auth_cap=1, random etc set),
>>       replies w/ temp asoc with INIT_ACK
>> 4) asoc B gets INIT_ACK, calls sctp_sf_do_5_1C_ack (and thus SCTP_PEER_INIT
>>     via interpreter), sees auth_cap=1, stores random etc; asoc B transitions
>>     from COOKIE_WAIT into COOKIE_ECHOED
>> 5) asoc A calls into sctp_sf_do_5_2_4_dupcook(), does the tietag compare,
>>     finds action B, creates temp asoc calls sctp_process_init() on it
>>     sees auth_cap=1, random etc; then we call into sctp_assoc_update()
>>     and migrate all params; what I see there is that random, chunks, hmac
>>     migrate from NULL each to the new values stored in the temp asoc
>>     (and thus we'd need auth_cap as well to be correct); after that, I
>>     see that asoc A goes from COOKIE_WAIT into ESTABLISHED (which seems
>>     to be in accordance to the RFC: "The endpoint should stay in or enter
>>     the ESTABLISHED state but it MUST ...")
>
> I see.
>
>> 6) later on, asoc B goes from COOKIE_ECHOED into ESTABLISHED
>>
>> So that led me to the resolution of transferring 'caps' over via
>> sctp_assoc_update(). In that case, asoc A transitions from 0 -> 1
>> as previous 'caps' haven't been stored in the actual asoc. It stayed
>> so far always in a temp asoc that we threw away after a reply.
>
> Thanks for the analysis.   The collisions in COOKIE_WAIT state is definitely
> a hole and it looks like all capabilities need to be updated and we should
> probably do an audit to make sure we don't miss anything else.

Thanks, I'll look into it and will respin the patch.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists