Re: Netfilter owner match breakage

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

On Tue, 2013-08-27 at 13:24 -0700, Phil Oester wrote:
> In commit 90ba9b19 (tcp: tcp_make_synack() can use alloc_skb()), Eric changed
> the call to sock_wmalloc in tcp_make_synack to alloc_skb.  In doing so,
> the netfilter owner match lost its ability to block the SYNACK packet on
> outbound listening sockets.  For example:
> 
> [phil@f19_main ~]$ id
> uid=1000(phil) gid=1000(phil) groups=1000(phil)
> 
> [phil@f19_main ~]$ sudo iptables -A OUTPUT -p tcp -m owner --uid-owner 1000 -j REJECT
> 

Oh well.

> [phil@f19_main ~]$ echo hi | nc -l -p 8888
> 
> Before this commit, attempting to connect to the port 8888 listener generated
> this:
> 
> 10.10.10.1.47457 > 10.10.10.110.8888: Flags [S], seq 1855270582 [...]
> 10.10.10.110 > 10.10.10.110: ICMP 10.10.10.1 tcp port 47457 unreachable, length 68
> 

Hmm... I think TCP stack should send more SYNACK, no ?

It sounds more logical to block the incoming SYN, but whatever.

> After this commit, the session is established but the first packet of the
> session gets rejected:
> 
> 10.10.10.1.47453 > 10.10.10.110.8888: Flags [S], seq 2089355862 [...]
> 10.10.10.110.8888 > 10.10.10.1.47453: Flags [S.], seq 2554257713, ack 2089355863 [...]
> 10.10.10.1.47453 > 10.10.10.110.8888: Flags [.], ack 1, win 115 [...]
> 10.10.10.110 > 10.10.10.110: ICMP 10.10.10.1 tcp port 47453 unreachable, length 63
> 
> Reverting 90ba9b19 (and adjusting for the removal of s_data_desired) fixes the
> problem.  Is there a better way to do this?

Well, a revert seems OK to me, it was not a critical change.