Re: [PATCH] Allow TCP connections to cache SYN packet for userspace inspection

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

On Fri, 2015-05-01 at 13:43 -0400, Eric B Munson wrote:
> In order to enable policy decisions in userspace, the data contained in
> the SYN packet would be useful for tracking or identifying connections.
> Only parts of this data are available to userspace after the hand shake
> is completed.  This patch exposes a new setsockopt() option that will,
> when used with a listening socket, ask the kernel to cache the skb
> holding the SYN packet for retrieval later.  The SYN skbs will not be
> saved while the kernel is in syn cookie mode.
> 
> The same option will ask the kernel for the packet headers when used
> with getsockopt() with the socket returned from accept().  The cached
> packet will only be available for the first getsockopt() call, the skb
> is consumed after the requested data is copied to userspace.  Subsequent
> calls will return -ENOENT.  Because of this behavior, getsockopt() will
> return -E2BIG if the caller supplied a buffer that is too small to hold
> the skb header.
> 
> Signed-off-by: Eric B Munson <emunson@akamai.com>
> Cc: Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
> Cc: James Morris <jmorris@namei.org>
> Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
> Cc: Patrick McHardy <kaber@trash.net>
> Cc: netdev@vger.kernel.org
> Cc: linux-api@vger.kernel.org
> Cc: linux-kernel@vger.kernel.org
> ---

We have a similar patch here at Google, but we do not hold one skb and
dst per saved syn. That can be ~4KB for some drivers.

Only a kmalloc() with the needed part (headers), usually less than 128
bytes. We store the length in first byte of this allocation.

This has a huge difference if you want to have ~4 million request socks.