Tor design proposals

Filename: 221-stop-using-create-fast.txt
Title: Stop using CREATE_FAST
Authors: Nick Mathewson
Created: 12 August 2013
Target: 0.2.5.x
Status: Closed

0. Summary

   I propose that in 0.2.5.x, Tor clients stop sending CREATE_FAST
   cells, and use CREATE or CREATE2 cells instead as appropriate.

1. Introduction

   The CREATE_FAST cell was created to avoid the performance hit of
   using the TAP handshake on a TLS session that already provided what
   TAP provided: authentication with RSA1024 and forward secrecy with
   DH1024.  But thanks to the introduction of the ntor onionskin
   handshake in Tor 0.2.4.x, for nodes with older versions of OpenSSL,
   the TLS handshake strength lags behind with the strength of the onion
   handshake, and the arguments against CREATE no longer apply.

   Similarly, it's good to have an argument for circuit security that
   survives possible breakdowns in TLS. But when CREATE_FAST is in use,
   this is impossible: we can only argue forward-secrecy at the first
   hop of each circuit by assuming that TLS has succeeded.

   So let's simply stop sending CREATE_FAST cells.

2. Proposed design

   Currently, only clients will send CREATE_FAST, and only when they
   have FastFirstHopPK set to its default value, 1.

   I propose that we change "FastFirstHopPK" from a boolean to also
   allow a new default "auto" value that tells Tor to take a value from
   the consensus.  I propose a new consensus parameter, "usecreatefast",
   default value taken to be 1.

   Once enough versions of Tor support this proposal, the authorities
   should set the value for "usecreatefast" to be 0.

   In the series after that (0.2.6.x?), the default value for
   "FastFirstHopPK" should be 0.

   (Note that CREATE_FAST must still be used in the case where a client
   has connected to a guard node or bridge without knowing any onion
   keys for it, and wants to fetch directory information from it.)

3. Alternative designs

   We might make some choices to preserve CREATE_FAST under some
   circumstances.  For example, we could say that CREATE_FAST is okay if
   we have a TLS connection with a cipher, public key, and ephemeral key
   algorithm of a given strength.

   We might try to trust the TLS handshake for authentication but not
   forward secrecy, and come up with a first-hop handshake that did a
   simple curve25519 diffie-hellman.

   We might use CREATE_FAST only whenever ntor is not available.

   I'm rejecting all of the above for complexity reasons.

   We might just change the default for FastFirstHopPK to 0 in
   0.2.5.x-alpha.  It would make early users of that alpha easy for
   their guards to distinguish.

4. Performance considerations

   This will increase the CPU requirements on guard nodes; their
   cpuworkers would be more heavily loaded as 0.2.5.x is more
   adopted.

   I believe that, if guards upgrade to 0.2.4.x as 0.2.5.x is under
   development, the commensurate benefits of ntor will outweigh the
   problems here.  This holds even more if we wind up with a better ntor
   implementation or replacement.

5. Considerations on client detection

   Right now, in a few places, Tor nodes assume that any connection on
   which they have received a CREATE_FAST cell is probably from a
   non-relay node, since relays never do that.  Implementing this
   proposal would make that signal unreliable.

   We should do this proposal anyway.  CREATE_FAST has never been a
   reliable signal, since "FastFirstHopPK 0" is easy enough to type, and
   the source code is easy enough to edit.  Proposal 163 and its
   successors have better ideas here anyway.