Re: tcp option added by driver

Re: tcp option added by driver

[Christoph Paasch]

Hello,

On 22/04/14 - 05:27:55, Madhan wrote:
> hi ,
> Is it possible to update the option field of the tcp header dynamically by the driver rather than the change in tcp/ip stack .
> i.e Can i add a driver to implement mptcp rather than building entire kernel again ?

in principle it seems possible to do this at a driver-level. Alcatel-Lucent
implemented something similar using netfilter-hooks:
https://open-innovation.alcatel-lucent.com/projects/mptcp-proxy

Regular TCP was running on top, and the netfilter hooks allowed to
rewrite the TCP/IP header.


However, I think that doing it that way will be huge implementation effort
(considering that there exists already an open-source implementation at 
multipath-tcp.org) and it might be hard to achieve good performance and make it
scale properly.

You basically have to reimplement a full TCP-stack at the driver as each
TCP-subflow has to behave as regular TCP (3-way handshake, retransmissions,...)


Cheers,
Christoph

Re: tcp option added by driver

[Christoph Paasch]

On 22/04/14 - 10:00:43, Rishi Madhan wrote:
> >You basically have to reimplement a full TCP-stack at the driver as each
> >TCP-subflow has to behave as regular TCP (3-way handshake, retransmissions,...)
> Can it be done like the usb modem driver ? Like adding the NCM or QMI or RNDIS header on the data field .
> Rather than changing the entire tcp header in the driver level .

I don't know exactly what the usb modem driver does, but for MPTCP it is
much more complex than just adding some bits to a segment.

> In case of mptcp  only the option field has to be changed . 

And the TCP/IP-header as you want to send your traffic via different paths. 

> And instead of sending the mptcp flags (such as mp_cap ,mp_join)
> inside option . Can it be done on the data ?

Then, this would be a whole new protocol, no more MPTCP. That's a different
story, and the idea of putting everything inside the data instead of the
TCP-option space has been widely discussed at the IETF, but got rejected
because of deadlock-issues with such an approach.

> The main reason why i am tryin to implement in this way is the effect of the MIDDLEBOX . 

What do you mean by the effect of the middlebox?


Btw., I'm not sure if lkml is the right place to discuss this. Maybe
move the discussion to mptcp-dev.


Cheers,
Christoph


> and also that server need not
> rebuild kernel (most of them dont want to also)  and can dynamically load/unload driver on need basis .
> 
> Regards,
> Madhan
> 
> On Tue, Apr 22, 2014 at 1:44 PM, Christoph Paasch <christoph.paasch@uclouvain.be<mailto:christoph.paasch@uclouvain.be>> wrote:
> Hello,
> 
> On 22/04/14 - 05:27:55, Madhan wrote:
> > hi ,
> > Is it possible to update the option field of the tcp header dynamically by the driver rather than the change in tcp/ip stack .
> > i.e Can i add a driver to implement mptcp rather than building entire kernel again ?
> 
> in principle it seems possible to do this at a driver-level. Alcatel-Lucent
> implemented something similar using netfilter-hooks:
> https://open-innovation.alcatel-lucent.com/projects/mptcp-proxy
> 
> Regular TCP was running on top, and the netfilter hooks allowed to
> rewrite the TCP/IP header.
> 
> 
> However, I think that doing it that way will be huge implementation effort
> (considering that there exists already an open-source implementation at
> multipath-tcp.org<http://multipath-tcp.org>) and it might be hard to achieve good performance and make it
> scale properly.
> 
> You basically have to reimplement a full TCP-stack at the driver as each
> TCP-subflow has to behave as regular TCP (3-way handshake, retransmissions,...)
> 
> 
> Cheers,
> Christoph
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org<mailto:majordomo@vger.kernel.org>
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>