Gen-ART review for draft-ietf-ipsecme-esp-null-heuristics-05.txt

["Spencer Dawkins" <[e-mail]>]

I have been selected as the General Area Review Team (Gen-ART) reviewer for
this draft (for background on Gen-ART, please see
http://www.alvestrand.no/ietf/gen/art/gen-art-FAQ.html)

Please resolve these comments along with any other Last Call comments you
may receive.

Document: draft-ietf-ipsecme-esp-null-heuristics-05.txt
Reviewer: Spencer Dawkins
Review Date: 2010-02-15 (sorry!)
IETF LC End Date: 2010-02-10
IESG Telechat date: (not known)
Summary:

1.2.  Terminology

   IPsec Flow

      An IPsec flow is a stream of packets sharing the same source IP,
      destination IP, protocol (ESP/AH) and SPI.  Strictly speaking, the
      source IP does not need to be as part of the flow identification,
      but as it can be there depending on the receiving implementation
      it is safer to assume it is always part of the flow
      identification.

Spencer (clarity): Last sentence is difficult to parse. My suggestion is to
use something like

An IPsec flow is a stream of packets sharing the same source IP, destination
IP, protocol (ESP/AH) and SPI.  Strictly speaking, the source IP does not
need to be as part of the flow identification, but it can be. For this
reason, it is safer to assume that the source IP is always part of the flow
indentification.

2.1.  AH

   The another problem is that in the new IPsec Architecture [RFC4301]

Spencer (clarity): "The second problem" or "Another Problem" here...

   the support for AH is now optional, meaning not all implementations
   support it.  ESP-NULL has been defined to be mandatory to implement
   by the Cryptographic Algorithm Implementation Requirements for
   Encapsulating Security Payload (ESP) document [RFC4835].

   AH has also quite complex processing rules compared to ESP when
   calculating the ICV, including things like zeroing out mutable
   fields, also as AH is not as widely used than ESP, the AH support is
   not as well tested in the interoperability events.

Spencer (clarity): I think there needs to be a semicolon or other break
before "also".

2.2.  Mandating by Policy

   Another easy way to solve this problem is to mandate the use of ESP-
   NULL with common parameters within an entire organization.  This
   either removes the need for heuristics (if no ESP encrypted traffic
   is allowed at all) or simplifies them considerably (only one set of
   parameters needs to be inspected, e.g. everybody in the organization
   who is using ESP-NULL must use HMAC-SHA-1-96 as their integrity
   algorithm).  This does not work unless one of a pair of communicating
   machines is not under the same administrative domain as the deep

Spencer (minor): I don't understand. I expected this to say "DOES work
unless". The text says that's the only situation where it fails!

   inspection engine.  (IPsec Security Associations must be satisfactory
   to all communicating parties, so only one communicating peer needs to
   have a sufficiently narrow policy.)  Also, such a solution might
   require some kind of centralized policy management to make sure
   everybody in an administrative domain uses the same policy.

Spencer (minor): Is it fair to point out that this type of heuristic will
make changing the common attribute value you're looking for more difficult?
If you decide to move away from HMAC-SHA-1-96, for instance...

3.  Description of Heuristics

   As described in section 7, UDP encapsulated ESP traffic may also have
   have NAPT applied to it, and so there is already a 5-tuple state in
   the stateful inspection gateway

Spencer (nit): missing period for this sentence.

4.  IPsec flows

   ESP is a stateful protocol, meaning there is state stored in the both

Spencer (nit): s/the both/both/

   end nodes of the ESP IPsec SA, and the state is identified by the
   pair of destination IP and SPI.  End nodes also often fix the source
   IP address in an SA unless the destination is a multicast group.
   Typically most (if not all) flows of interest to an intermediate
   device are unicast, so it is safer to assume the receiving node also
   uses a source address, and the intermediate device should therefore
   do the same.  In some cases this might cause extraneous cached ESP
   IPsec SA flows, but by using the source address two distinct flows
   will never be mixed.  For sites which heavily use multicast, such
   traffic is deterministically identifiable (224.0.0.0/4 for IPv4 and
   ff00::0/8 for IPv6), and an implementation can save the space of
   multiple cache entries for a multicast flow by checking the
   destination address first.

   There are several reasons why a single packet might not be enough to
   detect type of flow.  One of them is that the next header number was
   unknown, i.e. if heuristics do not know about the protocol for the
   packet, it cannot verify it has properly detected ESP-NULL
   parameters, even when the packet otherwise looks like ESP-NULL.  If
   the packet does not look like ESP-NULL at all, then encrypted ESP
   status can be returned quickly.  As ESP-NULL heuristics should know
   the same protocols as a deep inspection device, an unknown protocol
   should not be handled any differently than a cleartext instance of an
   unknown protocol if possible.

Spencer (minor): Are you saying that it might not be possible to handle the
two things the same way? I don't understand why. Prohibited by policy, sure,
and there may be other reasons to treat them differently, but I don't 
understand why this is "should" ...

6.  Special and Error Cases

   Each ESP-NULL flow should also keep statistics about how many packets
   have been detected as garbage by deep inspection, how many have
   passed checks, or how many have failed checks with policy violations
   (i.e. failed because actual inspection policy failures, not because

Spencer (clarity): s/because actual/because of actual/

   the packet looked like garbage).  If the number of garbage packets
   suddenly increases (e.g. most of the packets start to be look like
   garbage according to the deep inspection engine), it is possible the
   old ESP-NULL SA was replaced by an identical-SPI encrypting ESP SA.
   If both ends use random SPI generation, this is a very unlikely
   situation (1 in 2^32), but it is possible that some nodes reuse SPI
   numbers (e.g. a 32-bit memory address of the SA descriptor), thus
   this situation needs to be handled.

8.1.  ESP-NULL format



   The currently defined ESP authentication algorithms have 4 different
   lengths for the ICV field.  Most commonly used is 96 bits, and after
   that comes 128 bit ICV lengths.

Spencer (clarity): the second sentence in this paragraph is confusing, but I
think it's also unnecessary. I suggest dropping it... the next paragraph
replaces it nicely, anyway.

   Different ICV lengths for different algorithm:

       Algorithm                           ICV Length
       -------------------------------     ----------
       AUTH_HMAC_MD5_96                    96
       AUTH_HMAC_SHA1_96                   96
       AUTH_AES_XCBC_96                    96
       AUTH_AES_CMAC_96                    96
       AUTH_HMAC_SHA2_256_128              128
       AUTH_HMAC_SHA2_384_192              192
       AUTH_HMAC_SHA2_512_256              256

                                 Figure 2

   In addition to the ESP authentication algorithms listed above, there
   is also encryption algorithm ENCR_NULL_AUTH_AES_GMAC which does not
   provide confidentiality but provides authentication, just like ESP-
   NULL does.  This algorithm has ICV Length of 128 bits, and it also
   requires eight bytes of IV.

Spencer (clarity): I'd add this algorithm to the table, and remove the first
part of the sentence so that you're just describing one of the table
entries.

8.2.  Self Describing Padding Check

   At this point a maximum of 1.6% of packets remain, so the next header

Spencer (minor): If I'm following you, this isn't "1.6% of packets", it's
"1.6% of possible byte values", or something like that, right?

   number is inspected.  If the next header number is known (and
   supported) then the packet can be inspected based on the next header
   number.  If the next header number is unknown (i.e. not any of those
   with protocol checking support) the packet is marked "unsure",
   because there is no way to detect the IV length without inspecting
   the inner protocol payload.

8.3.  Protocol Checks

   The worst case scenario is when an end node starts up communication,
   i.e. it does not have any previous flows through the device.
   Heuristics will run on the first few packets received from the end
   node.  The later subsections mainly cover these bring up cases, as

Spencer (clarity): suggest s/bring up/start-up/, or something like that.

   they are the most difficult.

8.3.1.  TCP checks

   The most obvious field, TCP checksum, might not be usable, as it is
   possible that the packet has already transited a NAT box, thus the IP
   numbers used in the checksum are wrong, thus the checksum is wrong.

Spencer (minor): this isn't something I'm smart about, but would you expect
to see NAT boxes changing IP addresses and not fixing-up transport
checksums? That's begging for the receiver of these packets to discard them
based on checksum mismatches, isn't it? I know a NAT could be doing
anything, but that that seems short-sighted.

   One good method of detection is if a packet is dropped then the next
   packet will most likely be a retransmission of the previous packet.

Spencer (minor): is this true when you have a transmit window size greater
than one packet, so that more than one packet is outstanding? I agree with
the heuristic, but not with the statement that it's a "good method of
detection" - I don't think it will be triggered very often for web browsers,
or or TCP-based streaming media, or anything that's not stop-and-wait.

   Thus if two packets are received with the same source, and
   destination port numbers, and where sequence numbers are either same
   or right after each other, then it's likely a TCP packet has been
   correctly detected.

8.3.4.  SCTP checks

   SCTP chunks can be inspected to see if their lengths are consistent
   across the total length of the IP datagram, so long as TFC padding is
   not present.

Spencer (nit): could you expand "TFC" (this is the first usage in the 
document)?

9.  Security Considerations

   Using ESP-NULL or especially forcing using of it everywhere inside
   the enterprise can have increased risk of sending confidential
   information where eavesdroppers can see it.

Spencer (minor): I'm not arguing with this statement, I'm just confused by 
it. "Increased risk" compared to what? Saying that forbidding encrypted ESP 
makes it easier to eavesdrop doesn't seem profound - was that what you 
meant?