Re: [MPTCP] Separating MPTCP packet processing from TCP

[Mat Martineau <mathew.j.martineau at linux.intel.com>]

[-- Attachment #1: Type: text/plain, Size: 4161 bytes --]


On Tue, 16 May 2017, Rao Shoaib wrote:

>
>
> On 05/15/2017 09:59 PM, Christoph Paasch wrote:
>> Hello Rao,
>> 
>> On 15/05/17 - 18:16:36, Rao Shoaib wrote:
>>> I am thinking the cleanest implementation would be to separate MPTCP
>>> processing from TCP. On the receive side, when a packet arrives TCP only
>>> does TCP processing (That leaves the code untouched), the packet is than
>>> passed upto MPTCP which does MPTCP processing and can either process the
>>> packet, drop it or send a reset.
>>> 
>>> The current implementation is doing receive processing in TCP because it
>>> wants to validate the packet and accept it in TCP  or not -- but why ? 
>>> that
>>> seems to be an implementation choice.
>> you can look at mptcp_handle_options for the conditions upon which we 
>> stop processing the segment, when coming from tcp_validate_incoming. 
>> For example, when a MP_FASTCLOSE is received, the subflow gets killed 
>> with a TCP-RST.
>> 
>> The same holds when receiving a DSS-option without DSS-checksum 
>> although the DSS-checksum has been negotiated.
>> 
>> 
>> The question would be, whether we can delay the killing of the subflow &
>> sending of the RST until later. This would mean that the packet gets
>> completely handled (acks processed, incoming data acknowledged,...)
>> by the TCP-stack and later on the subflow gets killed.
>> This *might* be ok from a protocol-perspective, but I'm not sure.
> Hi Christoph,
>
> That is exactly what I am suggesting. I know this is fine from the protocol 
> perspective and I don't see why it can not be implemented. That will leave 
> the TCP code untouched on the receive side and make the upstream guys happy 
> because that is the fast path. Last night I was looking at the control path, 
> that needs to be cleaned up also but that is the slow path.

This looks like a good direction to me, with the disclaimer that I don't 
have the depth of MPTCP protocol expertise that Christoph does. I searched 
through the RFC for all of the "MUSTs", and didn't turn up anything that 
would require earlier processing of the received MPTCP options. Maybe the 
TCP layer will acknowledge the packet, but that should be fine for the 
MPTCP layer.

The cost of parsing the TCP options twice seems pretty low, since the 
first pass would skip over MPTCP and the second pass would skip over 
"regular" options.

>> 
>>> In the case where the receiver drops the packet in MPTCP, no data ack will
>>> be sent and MPTCP will re-transmit, It can retransmit even on the same 
>>> flow.
>>> To achieve this the code requires some change as the DSS option has to be
>>> saved. I think this is doable and is a much cleaner solution.
>> For incoming data, that's already how it is handled. We pass the segment to
>> the MPTCP-stack through the sk_data_ready callback, where we go over the
>> segments, check whether their DSS-checksum is correct, check if it is
>> in-window,... And if all is good, queue it at the MPTCP-level and send a
>> DATA_ACK.
> Yes. However the MPTCP options are parsed and mptcp_flags and dss offset are 
> extracted, no need to do that. Let MPTCP handle it. It's not a lot of 
> overhead and leaves the TCP code clean. This also removes the requirement of 
> trying to find space in  tcp_skb_cb to pass any information to MPTCP.
>
>> 
>>> Similarly we need to think about the Tx side -- That is not so straight
>>> forward but we need to think harder.

Agreed. One thing I'm thinking over is whether we can pre-populate some of 
the TCP header. We could set a bit in the control block to indicate that 
there is option data there already, and when the rest of the header is 
written the information could be read, moved, or replaced. One tricky part 
of this is making sure everything works right with TSO/GSO, etc.

>>> I can work on making the Rx changes but want to discuss it in case I am
>>> missing potential issues and if this is not a got option.

Are you approaching this as a refactor of the multipath-tcp.org kernel, or 
by building up from net-next?

Thanks,

--
Mat Martineau
Intel OTC