Re: [RFCv2 1/9] tcp: authopt: Initial support and key management

[David Ahern]

On 8/11/21 1:11 PM, Leonard Crestez wrote:
> On 11.08.2021 16:42, David Ahern wrote:
>> On 8/11/21 2:29 AM, Leonard Crestez wrote:
>>> On 8/10/21 11:41 PM, Dmitry Safonov wrote:
>>>> Hi Leonard,
>>>>
>>>> On Tue, 10 Aug 2021 at 02:50, Leonard Crestez <cdleonard@xxxxxxxxx>
>>>> wrote:
>>>> [..]
>>>>> +/* Representation of a Master Key Tuple as per RFC5925 */
>>>>> +struct tcp_authopt_key_info {
>>>>> +       struct hlist_node node;
>>>>> +       /* Local identifier */
>>>>> +       u32 local_id;
>>>>
>>>> There is no local_id in RFC5925, what's that?
>>>> An MKT is identified by (send_id, recv_id), together with
>>>> (src_addr/src_port, dst_addr/dst_port).
>>>> Why introducing something new to already complicated RFC?
>>>
>>> It was there to simplify user interface and initial implementation.
>>>
>>> But it seems that BGP listeners already needs to support multiple
>>> keychains for different peers so identifying the key by (send_id,
>>> recv_id, binding) is easier for userspace to work with. Otherwise they
>>> need to create their own local_id mapping internally.
>>>
>>
>> any proposed simplification needs to be well explained and how it
>> relates to the RFC spec.
> 
> The local_id only exists between userspace and kernel so it's not really
> covered by the RFC.

My point is that you need to document the uapi (however it ends up) and
how it is used to achieve the intent of the RFC.

> 
> There are objections to this and it seems to be unhelpful for userspace
> zo I will replace it with match by binding.
> 
> BTW: another somewhat dubious simplification is that I offloaded the RFC
> requirement to never add overlapping keys to userspace. So if userspace
> adds keys with same recvid that match the same TCP 4-tuple then
> connections will just start failing.
> 
> It's arguably fine to allow userspace misconfiguration to cause failures.
> 
> -- 
> Regards,
> Leonard