Re: [PATCH] sctp: Do not trigger BUG_ON when deleting assoc without primary path

[Daniel Borkmann <dborkman@redhat.com>]

On 10/17/2013 08:35 PM, Vlad Yasevich wrote:
> On 10/17/2013 02:25 PM, Daniel Borkmann wrote:
>> On 10/17/2013 08:01 PM, Daniel Borkmann wrote:
>>> On 10/17/2013 07:30 PM, Vlad Yasevich wrote:
>>>> It is possible to enter sctp_cmd_delete_tcb() without having a
>>>> primary path.  The situations this most often happens in is
>>>> when duplication cookie processing is triggered.  In this
>>>> case, we are deleting a temporarily created association that
>>>> is not fully populated.   Additially, at the time we
>>>> are deleting the offending association, it is really too
>>>> late to issue a BUG!
>>>>
>>>> This was introduced by:
>>>> commit f9e42b853523cda0732022c2e0473c183f7aec65
>>>>     net: sctp: sideeffect: throw BUG if primary_path is NULL
>>>
>>> Sure, lets remove it, but then we could still get a WARN() [sure,
>>> better than BUG], if the user at the very same time checks procfs
>>> through sctp_seq_dump_local_addrs(), see discussion we had here [1]:
>>>
>>>   It may trigger the crash later if the user performs some action on the
>>>   association that touches the primary. That's the reason why I was
>>>   proposing the checks below.
>>>
>>>   With the checks in command interpreter, we are only left with the
>>>   possibility that primary_path changes to NULL during the association
>>>   lifetime, which code audit doesn't support right now.  If that ever
>>>   changes we would at least have a bit more information to go on.
>>>
>>>   [1] http://patchwork.ozlabs.org/patch/251099/
>>
>> Meaning, all I'm saying is that with f9e42b853 we wanted to find exactly
>> such a case we have right now, that is, that an assoc could enter the
>> hashtable w/o primary path, no?
>
> But it didn't enter a hash table in this case.  SCTP_CMD_NEW_ASOC
> was never issued.  The sequence was:
>      SCTP_CMD_SET_ASOC
>      SCTP_CMD_DELETE_TCB
>
> Such association would never be found through /proc since it was never
> hashed.  Such association would never be found the user since it
> is only really alive while the packet is processed.  By all rights
> it should be marked as 'temp', but it isn't due to cookie processing.
>
> May be we should update cookie processing function to allow it
> to create temp associations if so desired.

Yes, I think that might be the better way to move on.

> -vlad