ICMP attacks against TCP and PMTUD

[Nikolay Denev]

On Jan 13, 2012, at 11:47 AM, Andre Oppermann wrote:

> On 12.01.2012 18:55, Nikolay Denev wrote:
>> Hello,
>> 
>> A web server that I administer running Nginx and FreeBSD-7.3-STABLE was recently
>> under a ICMP attack that generated a large amount of outgoing TCP traffic.
>> With some tcpdump and netflow analysis it was evident that the attachers are using
>> ICMP host-unreach need-frag messages to make the web server
>> retransmit multiple times, giving a amplification factor of about 1.6.
>> Then I noticed RFC5927 ( http://www.faqs.org/rfcs/rfc5927.html ) and specifically section 7.2
>> which discusses countermeasures against such attacks. The text reads :
>> 
>>    This section describes a modification to the PMTUD mechanism
>>    specified in [RFC1191] and [RFC1981] that has been incorporated in
>>    OpenBSD and NetBSD (since 2005) to improve TCP's resistance to the
>>    blind performance-degrading attack described in Section 7.1.  The
>>    described counter-measure basically disregards ICMP messages when a
>>    connection makes progress, without violating any of the requirements
>>    stated in [RFC1191] and [RFC1981].
>> 
>> The RFC is recent (dated from July 2010), and it mentions several times Linux, Free,Open and NetBSD,
>> but exactly in this paragraph it is mentioning only Net and OpenBSD's, thus I'm asking if
>> anyone has idea if these modifications were being put into FreeBSD?
> 
> We haven't implemented this (yet).
> 
>> I quickly glanced upon the source, but the TCP code is a bit too much for me :)
>> 
>> Also if anybody has observed similar attack, how are you protecting yourself from it?
>> Simply blocking host-unreach need-frag would break PMTUD.
> 
> We have a sysctl called "net.inet.tcp.minmss" which lower-bounds the
> MSS we accept in SYN and ICMP need frag messages.  It defaults to 216
> as 256 is the smallest allowable MTU in the Internet.  The only known
> user of MTU 256 is packet radio which isn't exactly much used on the
> common Internet.  You should be able to safely increase this value to
> 536.  If you are willing to live with a little bit of fall-out then
> 1220 is a good value as well.
> 
>> P.S.: I know 7.3 is pretty old, and I've planned upgrade to 8.2. I'm also curious if 8.2 will behave differently.
> 
> No.
> 
> -- 
> Andre

Thanks for the info Andre.

I'm now looking again at the pcap and I'm a bit confused.
First the possible attacker sends the ICMP need-frag packets with "MTU of next hop" set to zero,
which in 2012 shouldn't be very common?
Then when my server sends 66 byte FIN/ACK packet,
the attacker continues to send need-frag ICMPs and the FreeBSD host sends again
FIN/ACK packets.
Later on he sends again ICMP need-frag packets, but with size of about 1048 bytes,
with very large part of the original packets payload, instead of the required several bytes,
this then triggers excessive retransmits from the FreeBSD host which generates a lot of traffic.
The retransmits are roughly ~300-500 byte packets.