torspec/proposals/186-multiple-orports.txt

Filename: 186-multiple-orports.txt
Title: Multiple addresses for one OR or bridge
Author: Nick Mathewson
Created: 19-Sep-2011
Supersedes: 118
Status: Closed
Target: 0.2.4.x+

Status:

  This proposal is partially implemented to the extent needed to allow nodes
  to have one IPv4 and one IPv6 address.

Overview:

  This document is a proposal for servers to advertise multiple
  address/port combinations for their ORPort.

  It supersedes proposal 118.

Motivation:

  Sometimes servers want to support multiple ports for incoming
  connections, either in order to support multiple address families
  (ie, to add IPv6 support), to better use multiple interfaces, or
  to support a variety of FascistFirewallPorts settings.  This is
  easy to set up now, but there's no way to advertise it to clients.

Configuring additional addresses and ports:

  In consonance with our changes to the (Socks|Trans|NATD|DNS)Port
  options made in 0.2.3.x for proposal 171, I make a corresponding
  change to allow multiple ORPort options and deprecate
  ORListenAddress.

  The new syntax will be:

      "ORPort" PortDescription Option*

      Option = "NoAdvertise" | "NoListen" | "AllAddrs" | "IPV4Only"
          | "IPV6Only"

      PortDescription = PORTLIST |
                        ADDRESS ":" PORTLIST |
                        Hostname ":" PORTLIST

      (PORTLIST and ADDRESS are defined below.)

  The 'NoAdvertise' option performs the function of the old
  ORListenAddress option.  If it is set, we bind a port, but
  don't put it in our descriptor.

  The 'NoListen' option tells Tor to advertise an address, but not
  bind to it.  The operator needs to use some other mechanism to
  ensure that ports are redirected to ports that _are_ listened on.

  The 'AllAddrs' option tells Tor that if no address is given in the
  PortDescription part, we should bind/advertise every one of our
  publicly visible unicast addresses; and that if a hostname address
  is given in the PortDescription, we should bind/advertise every
  publicly visible unicast address that the hostname resolves to.
  (Q: Should this be on by default?)   The 'IPv4Only' and 'IPv6Only'
  options tell Tor to interpret such situations as applying only to
  IPv4 addresses or to IPv6 addresses.

  As with the client *Port options, only the old format or the new
  format are allowed: either a single numeric ORPort and zero or
  more ORListenAddress options, or a set of one or more
  ORPorts in the new extended format.

  In current operating systems (unless we get into crazy nonportable
  tricks) we need to use one socket for every address:port that Tor
  binds on.  As a sanity check, we can limit the number of such sockets
  we use to, say, something between 8 and 64.  If you want to bind lots
  of address:port combinations, you'll want to do it at the
  firewall/routing level.

  Example: We want to bind on 0.0.0.0:9001

     ORPort 9001

  Example: Our firewall is redirecting ports 80, 443, and 7000
  on all hosts in 18.244.2.0 onto our port 2929.

     ORPort 2929 noadvertise
     ORPort 18.244.2.0:80,443,7000 nolisten

  Example: We have a dynamic DNS provider that maps
  tornode.example.com to our current external IPv4 and IPv6
  addresses.  Our firewall forwards port 443 on those addresses to our
  port 1337.

     ORPort 1337 noadvertise alladdrs
     ORPort tornode.example.com:443 nobind alladdrs

Self-testing:

  Right now, Tor nodes need to check every port that they advertise
  before they declare themselves reachable.  If a Tor has
  a lot of advertised ports, that could be prohibitive.
  Instead, it should try a sample of ports for each address.  It should
  not advertise any given ORPort line until it has tried
  extending to or connecting to a sample of the address/port
  combinations.

  It will now be possible for a Tor node to find that some addresses
  work and others do not.  In this case, the node should only advertise
  ORPort lines that have been checked.  (As a consequence, the node
  should not advertise any address unless at least one ORPort without
  nolisten has been specified.)

  {Until support is added for extend cells to IPv6 addresses, it
  will only be possible to test IPv6 addresses by connecting
  directly.  We might want to just skip self-testing those until we
  have IPv6 extend support.}

New descriptor syntax:

  We add a new line in the router descriptor, "or-address".  This line
  can occur zero, one, or multiple times.  Its format is:

      or-address SP ADDRESS ":" PORTLIST NL

      ADDRESS = IPV6ADDR | IPV4ADDR
      IPV6ADDR = an ipv6 address, surrounded by square brackets.
      IPV4ADDR = an ipv4 address, represented as a dotted quad.
      PORTLIST = PORTSPEC | PORTSPEC "," PORTLIST
      PORTSPEC = PORT
      PORT = a number between 1 and 65535 inclusive.

  [This is the regular format for specifying sets of addresses and
  ports in Tor.]

  A descriptor should not include an or-address line that does
  nothing but duplicate the address:port pair from its "router"
  line.

  A node must not list more than 8 or-address lines.

  A PORTLIST must have no more than 16 PORTSPEC entries, and its entries must
  be disjoint.

  (Q: Any reason to allow more than 2?  Multiple interfaces, I guess.)

New authority behavior:

  The same rationale applies as for self-testing.  An authority
  needs to test the main address:port from the router line, and
  every or-address line.  For or-address lines that contain
  multiple ports, it needs to test all of them if they are few, or a
  sample if they are not.

  An authority shouldn't list a node as Running unless every
  or-address line it advertises looks like it will work.

Consensus directories and microdescriptors:

  We introduce a new line type for microdescriptors and consensuses,
  "a".  Each "a" line has the same format as an or-address line.
  The "a" lines (if any) appear immediately after the "r" line for a
  router in the consensus, and immediately after the "onion-key"
  entry in a microdescriptor.

  Clients that use microdescriptors should consider a node's
  addresses to be the address:port listed in the "r" line of a
  consensus, plus all "a" lines for that node in the consensus, plus
  all "a" lines for that node in its microdescriptor.  Clients
  that use full descriptors should consider a node's addresses to be
  everything listed in its descriptor.

  We will have to define a new voting algorithm version; when using
  this version or later, votes should include a single "a" line for
  every relay that has an IPv6 address, to include the first IPv6
  line in its descriptor.  (If there are no IPv6 or-address lines, then
  they shouldn't include any "a" lines.)  The remaining or-address
  lines will turn into "a" lines in the microdescriptor.

  As with other data in the vote derived from the descriptor, the
  consensus will include whichever set of "a" lines are given by the
  most authorities who voted for the descriptor digest that will be
  used for the router.

Directory authorities with more addresses:

  We need a way for a client to configure a TrustedDirServer as
  having multiple OR addresses, specifically so that we can give at
  least one default authority an IPv6 address for bootstrapping
  purposes.

  (Q: Do any of the current authorities have stable IPv6 addresses?)

  We will want to allow the address in a "dir-source" line in a vote
  to contain an IPv6 address, and/or allow voters to list themselves
  with more addresses in votes/consensuses.  But right now, nothing
  actually uses the addresses listed for voters in dir-source lines
  for anything besides log messages.

Client behavior:

  I propose that initially we shouldn't change client behavior too
  much here.

  (Q: Is there any advantage to having a client choose a random
  address?  If so we can do it later.  If not, why list any more
  than one IPv4 and one IPv6 address?)

  Tor clients not running with bridges, and running with IPv4
  support, should still use the address and ORPort as advertised in
  the "router" or "r" line of the appropriate directory object.

  Tor clients not running with bridges, and running without IPv4
  support, should use the first listed IPv6 address for a node,
  using the lowest-numbered listed port for that address.  They
  should only connect to nodes with an IPv6 address.

  Clients should accept Bridge lines with IPv6 addresses, and
  address:port sets, in addition to the lines they currently accept.

  Clients, for now, should only use the address:port from the router
  line when making EXTEND cells; see below.

Nodes without IPv4 addresses:

  Currently Tor requires every node or bridge to have an IPv4
  address.  We will want to maintain this property for the
  foreseeable future, but we should define how a node without an IPv4
  address would advertise itself.

  Right now, there's no way to do that: if anything but an IPv4
  address appears in a router line of a routerdesc, or the "r" line of
  a consensus, then it won't parse.  If something that looks like an
  IPv4 address appears there, clients will (I believe) try to
  connect to it.

  We can make this work, though: let's allow nodes to list themselves
  with a magic IPv4 address (say, 127.1.1.1) if they have
  or-address entries containing only IPv6 address.  We could give
  these nodes a new flag other than Running to indicate that they're
  up, and not give them the Running flag.  That way, old clients
  would never try to use them, but new clients could know to treat
  the new flag as indicating that the node is running, and know not
  to connect to a node listed with address 127.1.1.1.

Interaction with EXTEND and NETINFO:

  Currently, EXTEND cells only support IPv4 addresses, so we should
  use only those.  There is a proposal draft to support more address
  types.

  A server's NETINFO cells must list all configured addresses for a
  server.

Why not extend DirPort this way too?

  Because clients are all using BEGINDIR these days.

  That is, clients tunnel their directory requests inside OR
  connections, and don't generally connect to DirPorts at all.

Why not have address and port ranges?

  Earlier drafts of this proposal suggested that servers should provide
  ranges of addresses, specified with bitmasks.  That's a neat idea for
  circumvention, but if we did that, you wouldn't want to advertise
  publicly that you have an entire address range.

  Port ranges are out because I don't think they would actually get used
  much, and they add a fair bit of complexity.

Coding impact:

  In addition to the obvious changes, we need to audit everything
  that looks up or compares OR connections and nodes by address:port
  under the assumptions that each node has only a single address or
  ORPort.

TODO:

  * Make it so that authorities can vote on which addresses are working
    somehow.

  * Specify some way to say "I only want to connect to v4/v6 addresses".

  * Come up with a better alternative to running6 for the longterm?