[tor-dev] DNS(SEC) draft update

[Ondrej Mikle]

Hi Nick,

I've revised the DNS draft, attaching it. In section 4 there are some options
for integration with libunbound, but each of them requires some work with the
stock libunbound code.

Ondrej
-------------- next part --------------
Filename: xxx-dns-dnssec.txt
Title: Support for full DNS and DNSSEC resolution in Tor
Authors: Ondrej Mikle
Created: 4 February 2012
Modified: 19 August 2012
Status: Draft

0. Overview

  Adding support for any DNS query type to Tor, as well as DNSSEC support.

0.1. Motivation

  Many applications running over Tor need more than just resolving FQDN to
  IPv4 and vice versa. Sometimes to prevent DNS leaks the applications have to
  be hacked around to be supplied necessary data by hand (e.g. SRV records in
  XMPP). TLS connections will benefit from planned TLSA record that provides
  certificate pinning to avoid another Diginotar-like fiasco.
  
  DNSSEC is part of the DNS protocol and the most appropriate place for DNSSEC
  API would be probably in OS libraries (e.g. libc). However that will
  probably take time until it becomes widespread.

  On the Tor's side (as opposed to application's side), DNSSEC will provide
  protection against DNS cache-poisoning attacks (provided that exit is not
  malicious itself, but still reduces attack surface).

1. Design

1.1 New cells

  There will be two new cells, RELAY_DNS_BEGIN and RELAY_DNS_RESPONSE (we'll
  use DNS_BEGIN and DNS_RESPONSE for short below).

  DNS_BEGIN payload:

    DNS packet data (variable length)

  The DNS packet must be generated internally by libunbound to avoid
  fingerprinting users by differences in client resolvers' behavior.

  DNS_RESPONSE payload:
  
    total length (2 octets)
    data         (variable)
  
  Data contains the reply DNS packet or its part if packet would not fit into
  the cell. Total length describes length of complete response packet, thus
  one DNS_BEGIN may be answered by multiple DNS_RESPONSE cells.

  DNS_BEGIN must use a non-zero, distinct StreamID, corresponding DNS_RESPONSE
  will use the same StreamID. Similarly to RELAY_RESOLVE(D), no actual stream
  is created.

  AXFR and IXRF are not supported in this cell by design (see specialized tool
  below).

2. Interfaces to applications

  DNSPort evdns - existing implementation will be updated to use DNS_BEGIN.
  
3. Limitations on DNS query

  Query class is limited to IN (INTERNET) since the only other useful class
  CHAOS is practical for directly querying authoritative servers (OR in this
  case acts as a recursive resolver). Query for class other than IN will
  return REFUSED in the inner DNS packet.

  Multiple questions in a single packet are not supported and OR will respond
  with REFUSED as the DNS error code.

  All query RR types are allowed.

  [XXXX I originally thought about some exit policy like "basic RR types" and
  "all RRs", but managing such list in deployed nodes with extra directory
  flags outweighs the benefit. Maybe disallow ANY RR type? ]

  Client as well as OR MUST block attempts to resolve local RFC 1918, 4193,
  4291 adresses (PTR). REFUSED will be returned as DNS error code from OR.

  Request for special names (.onion, .exit, .noconnect) will return REFUSED.

4. Implementation notes

  Client will periodically purge incomplete DNS replies. Any unexpected
  DNS_RESPONSE will be dropped.

  AD flag must be zeroed out on client unless validation is performed.

  [XXXX libunbound lowlevel API, Tor+libunbound libevent loop

  libunbound doesn't publicly expose all the necessary parts of low-level API.
  It can return the received DNS packet, but not let you construct a packet
  and get it in wire-format, for example.

  Options I see:

  a) patch libunbound to be able feed wire-format DNS packets and add API to
  obtain constructed packets instead of sending over network

  b) replace bufferevents for sockets in unbound with something like
  libevent's paired bufferevents. This means that data extracted from
  DNS_RESPONSE/DNS_BEGIN cells would be fed directly to some evbuffers that
  would be picked up by libunbound. It could possibly result in avoiding
  background thread of libunbound's ub_resolve_async running separate libevent
  loop.

  c) bind to some arbitrary local address like 127.1.2.3:53 and use it as
  forwarder for libunbound. The code there would pack/unpack the DNS packets
  from/to libunbound into DNS_BEGIN/DNS_RESPONSE cells. It wouldn't require
  modification of libunbound code, but it's not pretty either. Also the bind
  port must be 53 which usually requires superuser privileges.

  Code of libunbound is fairly complex for me to see outright what would the
  best approach be.
  ]

5. Separate tool for AXFR

  The AXFR tool will have similar interface like tor-resolve, but will
  return raw DNS data. 
  
  Parameters are: query domain, server IP of authoritative DNS. 
  
  The tool will transfer the data through "ordinary" tunnel using RELAY_BEGIN
  and related cells.
  
  This design decision serves two goals:

  - DNS_BEGIN and DNS_RESPONSE will be simpler to implement (lower chance of
    bugs)
  - in practice it's often useful do AXFR queries on secondary authoritative
    DNS servers

  IXFR will not be supported (infrequent corner case, can be done by manual
  tunnel creation over Tor if truly necessary).

6. Security implications

  Transaction ID is provided randomly by libunbound, no need to modify.

  As proposal 171 mentions, we need mitigate circuit correlation. One solution
  would be keeping multiple streams to multiple exit nodes and picking one at
  random for DNS resolution. Other would be keeping DNS-resolving circuit open
  only for a short time (e.g. 1-2 minutes). Randomly changing the circuits
  however means that it would probably incur additional latency since there
  would likely be a few cache misses on the newly selected exits.
  

7. TTL normalization idea 

  A bit complex on implementation, because it requires parsing DNS packets at
  exit node.

  TTL in reply DNS packet MUST be normalized at exit node so that client won't
  learn what other clients queried. The normalization is done in following
  way:

  - for a RR, the original TTL value received from authoritative DNS server
    should be used when sending DNS_RESPONSE, trimming the values to interval
    [5, 600]
  - does not pose "ghost-cache-attack", since once RR is flushed from
    libunbound's cache, it must be fetched anew