Re: [PATCH 2/3] tcp: Add ESP encapsulation support

[Eric Dumazet <eric.dumazet@gmail.com>]

On Fri, 2018-01-12 at 00:21 +1100, Herbert Xu wrote:
> This patch adds the plumbing in TCP for ESP encapsulation support
> per RFC8229.
> 
> The patch mostly deals with inbound processing, as well as enabling
> TCP encapsulation on a socket through setsockopt.  The outbound
> processing is dealt with in the ESP code as is done for UDP.
> 
> The inbound processing is split into two halves.  First of all,
> the softirq path directly intercepts ESP packets and feeds them
> into the IPsec stack.  Most of the time the packet will be freed
> right away if it contains complete ESP packets.  However, if
> the message is incomplete or it contains non-ESP data, then the
> skb will be added to the receive queue.  We also add packets to
> the receive queue if it is currently non-emtpy, in order to
> preserve sequence number continuity and minimise the changes
> to the TCP code.
> 
> On the user-space facing side, packets marked as ESP-only are
> skipped and not visible to user-space.  However, some ESP data
> may seep through.  For example, if we receive a partial message
> then we will always give it to user-space regardless of whether
> it turns out to be ESP or not.  So user-space should be prepared
> to skip ESP messages (SPI != 0).
> 
> There is a little bit of code dealing with the encapsulation side.
> In particular, if encapsulation data comes in while the socket
> is owned by user-space, the packets will be stored in tp->encap_out
> and processed during release_sock.
> 
> Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
> ---
> 
>  include/linux/tcp.h      |   15 ++
>  include/net/tcp.h        |   27 +++
>  include/uapi/linux/tcp.h |    1 
>  include/uapi/linux/udp.h |    1 
>  net/ipv4/tcp.c           |   68 +++++++++
>  net/ipv4/tcp_input.c     |  326 +++++++++++++++++++++++++++++++++++++++++++++--
>  net/ipv4/tcp_ipv4.c      |    1 
>  net/ipv4/tcp_output.c    |   48 ++++++
>  8 files changed, 473 insertions(+), 14 deletions(-)
> 

Ouch...

Is there any chance this can be done with almost no change in TCP
stack, using a layer model ? ( net/kcm comes to mind )

NFS uses TCP sockets, but does not invade TCP stack either.

I believe Christoph Paasch sent a patch series during holidays trying
to cleanup the MD5 mess (I had no time reviewing it, sorry)