r9776@totoro: nickm | 2007-01-18 14:37:01 -0500

[tor.git] / doc / TODO

$Id$
Legend:
SPEC!!  - Not specified
SPEC    - Spec not finalized
N       - nick claims
R       - arma claims
P       - phobos claims
        - Not done
        * Top priority
        . Partially done
        o Done
        d Deferrable
        D Deferred
        X Abandoned

    X <nickm> "Let's try to find a way to make it run and make the version
       match, but if not, let's just make it run."
    X <arma> "should we detect if we have a --with-ssl-dir and try the -R
      by default, if it works?"

Items for 0.1.2.x, real soon now:
? - Bug: combination of things:
    When we've been idle a long time, we stop fetching server
    descriptors. When we then get a socks request, we build circuits
    immediately using whatever descriptors we have, rather than waiting
    until we've fetched correct ones.

N - Test guard unreachable logic; make sure that we actually attempt to
    connect to guards that we think are unreachable from time to time.
    Make sure that we don't freak out when the network is down.

R - Reconstruct ChangeLog; put rolled-up info in ReleaseNotes or something.

Items for 0.1.2.x:
N - enumerate events of important things that occur in tor, so vidalia can
    react.
    o Backend implementation
    o Actually list all the events (notice and warn log messages are a good
      place to look.)  Divide messages into categories, perhaps.
    o Specify general event system
    o Specify actual events.
    - Implement or defer remaining events
    - Implement or defer GETINFO list of current status events.
    - Clean up relevant bits of control-spec.txt

  . Have (and document) a BEGIN_DIR relay cell that means "Connect to your
    directory port."
    o Implement
R   - turn the received socks addr:port into a digest for setting .exit
    - be able to connect without having a server descriptor, to bootstrap.
R     - handle connect-dir streams that don't have a chosen_exit_name set.
      o include ORPort in DirServers lines so we can know where to connect.
        list the orport as 0 if it can't handle begin_dir.
        - List orports of actual dirservers..

  - Servers are easy to setup and run: being a relay is about as easy as
    being a client.
    . Reduce resource load
      o A way to alert controller when router flags change.
        o Specify: SETEVENTS NS
        o Implement
R       - Hunt for places that change networkstatus info that I might have
          missed.

  . option to dl directory info via tor:
    TunnelDirConns and PreferTunneledDirConns
R   - actually cause the directory.c functions to know about or_port
      and use it when we're supposed to.
N   - for tunneled edge conns, stop reading to the bridge connection
      when the or_conn we're writing to has a full outbuf.
      - This is iffy.  Really, it would've been better to stop writing
        on the dir conn when the edge conn's inbuf is getting full.
        But the issue there is that we package from edge
        connections aggressively until we hit their package windows or the
        circuit package windows, even if the buffer on the corresponding OR
        connection is pretty damn big.  This sucks from a RAM usage POV.
        Now, we could try to stop reading on the edges (or just the edges
        connected to a local bridge) when an or_conn's outbuf is full.  But
        if we're a server, and we stop reading on some exit conns when OR
        conns are full, soon OR conns will contain only traffic from other OR
        conns, and the exit data in question  will never get written.

      - The right solution in the long run (0.2.0) is probably as follows:
        - Remove socketpair-based bridges: use shared (or connected) buffers
          for communication, rather than sockets.
        - When relaying cells from an OR conn to an OR conn, have them wait
          in a queue on the or_circuit_t object; don't move them onto the
          target conn until there is space in the target conn's outbuf.
          Also, only package data from exitconns when there is space in the
          target conn's outbuf.
          - As an added advantage, this would let us kill stalled _circuits_
            when their buffers get too full, rather than killing entire OR
            conns.  But we must think about anonymity implications of that.
        - We'll probably want to do some kind of big refactoring of our
          dataflow when we do these changes; stuff is hairy enough already,
          and it will only get harrier with this stuff.

      - For the short run, our options seem to be:
        - Disable BEGIN_DIR support at the server-side unless it's
          explicitly turned on.
        - Go ahead and make directory bridge data not get produced when the
          corresponding or_conn is full, and accept the sometimes directory
          data will just never get written.
        - Rate-limit directory bridge data somehow when the corresponding
          or_conn is full.  Possibly based on the flush rate of the or_conn?
        - Reject BEGIN_DIR requests that we think will be insanely big.


N - DNS improvements
    . Asynchronous DNS
      - Make evdns use windows strerror equivalents.
      - Make sure patches get into libevent.
      - Verify that it works well on windows
    - Debug and re-enable server-side reverse DNS caching

  - Critical but minor bugs, backport candidates.
    - support dir 503s better
      o clients don't log as loudly when they receive them
      o they don't count toward the 3-strikes rule
        D But eventually, we give up after getting a lot of 503s.
        D Delay when we get a lot of 503s, rather than punting onto the
          servers that have given us 503s?
        o Add a 'BadDirectory' flag to statuses.
      o authorities should *never* 503 a cache, and should never 503
        network status requests.
      D They can 503 client descriptor requests when they feel like it.
        How can they distinguish? Not implemented for now, maybe
        should abandon.
      - update dir-spec with what we decided for each of these

  o Have a mode that doesn't write to disk much, so we can run Tor on
    flash memory (e.g. Linksys routers or USB keys).
    o Add AvoidDiskWrites config option.
    o only write state file when it's "changed"
      o crank up the numbers if avoiddiskwrites is on.
      D some things may not want to get written at all.
    o stop writing fingerprint every restart
    D more?

NR. Write path-spec.txt

  - Polishing
    - Profile client and server; fix slow spots
    - Address XXX012 items

  - Packaging
    - Tell people about OSX Uninstaller
    - Quietly document NT Service options
    - Switch canonical win32 compiler to mingw.
NR  D Get some kind of "meta signing key" to be used solely to sign
      releases/to certify releases when signed by the right people/
      to certify sign the right people's keys?  Also use this to cert the SSL
      key, etc.
    - If we haven't replaced privoxy, lock down its configuration in all
      packages, as documented in tor-doc-unix.html

  - Docs
    - More prominently, we should have a recommended apps list.
      - recommend gaim.
      - unrecommend IE because of ftp:// bug.
N   - torrc.complete.in needs attention?
N   - we should add a preamble to tor-design saying it's out of date.
N   - Document transport and natdport

  - Improvements to bandwidth counting
R   - look into "uncounting" bytes spent on local connections, so
      we can bandwidthrate but still have fast downloads.
R   - "bandwidth classes", for incoming vs initiated-here conns,
      and to give dir conns lower priority.
    . Write limiting; separate token bucket for write
      o preemptively give a 503 to some v1 dir requests
      o preemptively give a 503 to some v2 dir requests
        o Write function to estimate bytes needed for N descriptors
          statuses
      D per-conn write buckets
      D separate config options for read vs write limiting
        (It's hard to support read > write, since we need better
         congestion control to avoid overfull buffers there.  So,
         defer the whole thing.)

  - Forward compatibility fixes
    - Caches should start trying to cache consensus docs?
NR    - Design
N     - Implement, if we think it's smart.
    - Start uploading short and long descriptors; authorities should support
      URLs to retrieve long descriptors, and should discard short descriptors
      for now.  Later, once tools use the "long descriptor" URLs, authorities
      will serve the short descriptors every time they're asked for
      a descriptor.
NR    - Design
N     - Implement, if we think it's smart.

Topics to think about during 0.1.2.x development:
  * Figure out incentives.
    - (How can we make this tolerant of a bad v0?)
  * Figure out non-clique.
  * Figure out China.
  - Figure out partial network knowledge.
  - Figure out hidden services.
  - Design next-version protocol for directories
  - Design next-version protocol for connections

Deferred from 0.1.2.x:
P - Figure out why dll's compiled in mingw don't work right in WinXP.
P - Figure out why openssl 0.9.8d "make test" fails at sha256t test.
  - Directory guards
  - RAM use in directory authorities.
  - Memory use improvements:
    - Look into pulling serverdescs off buffers as they arrive.
    - Save and mmap v1 directories, and networkstatus docs; store them
      zipped, not uncompressed.
      - Switch cached_router_t to use mmap.
      - What to do about reference counts on windows?  (On Unix, this is
        easy: unlink works fine.  (Right?)  On Windows, I have doubts.  Do we
        need to keep multiple files?)
      - What do we do about the fact that people can't read zlib-
        compressed files manually?

  - If the client's clock is too far in the past, it will drop (or
    just not try to get) descriptors, so it'll never build circuits.
  - Tolerate clock skew on bridge relays.

  - A more efficient dir protocol.
    - Authorities should fetch the network-statuses amongst each
      other, consensus them, and advertise a communal network-status.
      This is not so much for safety/complexity as it is to reduce
      bandwidth requirements for Alice.
      - How does this interact with our goal of being able to choose
        your own dir authorities? I guess we're now assuming that all
        dir authorities know all the other authorities in their "group"?
    - Should we also look into a "delta since last network-status
      checkpoint" scheme, to reduce overhead further?
    - Extend the "r" line in network-status to give a set of buckets (say,
      comma-separated) for that router.
      - Buckets are deterministic based on IP address.
      - Then clients can choose a bucket (or set of buckets) to
        download and use.

  - Improvements to versioning.
    - When we connect to a Tor server, it sends back a cell listing
      the IP it believes it is using. Use this to block dvorak's attack.
      Also, this is a fine time to say what time you think it is.
      o Verify that a new cell type is okay with deployed codebase
      . Specify HELLO cells
      . Figure out v0 compatibility.
      - Implement

  - Eventdns improvements
    - Have a way to query for AAAA and A records simultaneously.
    - Improve request API: At the very least, add the ability to construct
      a more-or-less arbitrary request and get a response.
    - (Can we suppress cnames? Should we?)

  - Now that we're avoiding exits when picking non-exit positions,
    we need to consider how to pick nodes for internal circuits. If
    we avoid exits for all positions, we skew the load balancing. If
    we accept exits for all positions, we leak whether it's an internal
    circuit at every step. If we accept exits only at the last hop, we
    reintroduce Lasse's attacks from the Oakland paper.

  - We should ship with a list of stable dir mirrors -- they're not
    trusted like the authorities, but they'll provide more robustness
    and diversity for bootstrapping clients.

  - Simplify authority operation
    - Follow weasel's proposal, crossed with mixminion dir config format

  - A way to adjust router flags from the controller.
    (How do we prevent the authority from clobbering them soon after?)
  - a way to pick entry guards based wholly on extend_info equivalent;
    a way to export extend_info equivalent.

  - Better estimates in the directory of whether servers have good uptime
    (high expected time to failure) or good guard qualities (high
    fractional uptime).
    - AKA Track uptime as %-of-time-up, as well as time-since-last-down

  - Have a "Faster" status flag that means it. Fast2, Fast4, Fast8?
    - spec
    - implement

  - Failed rend desc fetches sometimes don't get retried. True/false?

  - Windows server usability
    - Solve the ENOBUFS problem.
      - make tor's use of openssl operate on buffers rather than sockets,
        so we can make use of libevent's buffer paradigm once it has one.
      - make tor's use of libevent tolerate either the socket or the
        buffer paradigm; includes unifying the functions in connect.c.
    - We need a getrlimit equivalent on Windows so we can reserve some
      file descriptors for saving files, etc. Otherwise we'll trigger
      asserts when we're out of file descriptors and crash.
M   - rewrite how libevent does select() on win32 so it's not so very slow.
      - Add overlapped IO

  - Add an option (related to AvoidDiskWrites) to disable directory caching.

Minor items for 0.1.2.x as time permits:
  - getinfo ns/name/moria2 doesn't include a "v" line, even when some
    network-statuses I have show it. I suppose the fix should go in
    networkstatus_getinfo_helper_single() in routerlist.c.
  - Unify autoconf search code for libevent and openssl.  Make code
    suggest platform-appropriate "devel" / "dev" / whatever packages
    if we can link but we can't find the headers.
  - Recommend polipo? Please?
  - Make documentation realize that location of system configuration file
    will depend on location of system defaults, and isn't always /etc/torrc.
  - Review torrc.sample to make it more discursive.
  o when reporting clock skew (both to logs and to controller), if it's
    taken 126 seconds to read from the directory, our clock skew estimate
    is 126 seconds wrong. use conn->timestamp_create or _lastwritten
    for a closer estimate?
  - a way to generate the website diagrams from source, so we can
    translate them as utf-8 text rather than with gimp.
R - add d64 and fp64 along-side d and fp so people can paste status
    entries into a url. since + is a valid base64 char, only allow one
    at a time. spec and then do.
  D don't do dns hijacking tests if we're reject *:* exit policy?
    (deferred until 0.1.1.x is less common)
  - When we export something from foo.c file for testing purposes only,
    make a foo_test.h file for test.c to include.
  - The Debian package now uses --verify-config when (re)starting,
    to distinguish configuration errors from other errors. Perhaps
    the RPM and other startup scripts should too?
  - add a "default.action" file to the tor/vidalia bundle so we can fix the
    https thing in the default configuration:
    http://wiki.noreply.org/noreply/TheOnionRouter/TorFAQ#PrivoxyWeirdSSLPort
  . Flesh out options_description array in src/or/config.c
  - Don't let 'newnym' be triggered more often than every n seconds.
  X If we try to publish as a nickname that's already claimed, should
    we append a number (or increment the number) and try again? This
    way people who read their logs can fix it as before, but people
    who don't read their logs will still offer Tor servers.
    - Fall back to unnamed; warn user; send controller event.  ("When we
      notice a 'Rejected: There is already a named server with this nickname'
      message... or maybe instead when we see in the networkstatuses that
      somebody else is Named with the name we want: warn the user, send a
      STATUS_SERVER message, and fall back to unnamed.")
! - Tor should bind its ports before dropping privs, so users don't
    have to do the ipchains dance.
  - Rate limit exit connections to a given destination -- this helps
    us play nice with websites when Tor users want to crawl them; it
    also introduces DoS opportunities.
  - Streamline how we pick entry nodes: Make choose_random_entry() have
    less magic and less control logic.
  - Christian Grothoff's attack of infinite-length circuit.
    the solution is to have a separate 'extend-data' cell type
    which is used for the first N data cells, and only
    extend-data cells can be extend requests.
    - Specify, including thought about anonymity implications.
  - Display the reasons in 'destroy' and 'truncated' cells under some
    circumstances?
  - We need a way for the authorities to declare that nodes are
    in a family.  Also, it kinda sucks that family declarations use O(N^2)
    space in the descriptors.
  - If the server is spewing complaints about raising your ulimit -n,
    we should add a note about this to the server descriptor so other
    people can notice too.
  - cpu fixes:
    - see if we should make use of truncate to retry
    X kill dns workers more slowly
  . Directory changes
    . Some back-out mechanism for auto-approval
      - a way of rolling back approvals to before a timestamp
        - Consider minion-like fingerprint file/log combination.
  - packaging and ui stuff:
    . multiple sample torrc files
    . figure out how to make nt service stuff work?
      . Document it.
    - Vet all pending installer patches
      - Win32 installer plus privoxy, sockscap/freecap, etc.
      - Vet win32 systray helper code

  - Improve controller
      - a NEWSTATUS event similar to NEWDESC.
      - change circuit status events to give more details, like purpose,
        whether they're internal, when they become dirty, when they become
        too dirty for further circuits, etc.
        - What do we want here, exactly?
        - Specify and implement it.
      - Change stream status events analogously.
        - What do we want here, exactly?
        - Specify and implement it.
      - Make other events "better".
      - Change stream status events analogously.
        - What do we want here, exactly?
        - Specify and implement it.
      - Make other events "better" analogously
        - What do we want here, exactly?
        - Specify and implement it.
      . Expose more information via getinfo:
        - import and export rendezvous descriptors
        - Review all static fields for additional candidates
      - Allow EXTENDCIRCUIT to unknown server.
      - We need some way to adjust server status, and to tell tor not to
        download directories/network-status, and a way to force a download.
      - Make everything work with hidden services
  - Directory system improvements
    - config option to publish what ports you listen on, beyond
      ORPort/DirPort.  It should support ranges and bit prefixes (?) too.
      - Parse this.
      - Relay this in networkstatus.

  - Be a DNS proxy.
    - Need a way to request address lookups (and allocate a stream ID for
      them) without having a corresponding client socket.
      - Once this is done, it would be nice to have a way to request address
        lookups from the controller without using SOCKS.
    - Refactor exit/entry side of DNS resolve: we don't need a connection_t;
      we can have an edge_connection_t and (say) dns_request_t both extend an
      edge_stream_t, and have p_streams and n_streams both be linked lists
      of edge_stream_t.

  - Look into generating torrc.{complete|sample}.in, tor.1.in,
    the HTML manual, and the online config documentation from a single
    source.

Future version:
  - Configuration format really wants sections.
  - Good RBL substitute.
  - Authorities should try using exits for http to connect to some URLS
    (specified in a configuration file, so as not to make the List Of Things
    Not To Censor completely obvious) and ask them for results.  Exits that
    don't give good answers should have the BadExit flag set.
  - Our current approach to block attempts to use Tor as a single-hop proxy
    is pretty lame; we should get a better one.
  . Update the hidden service stuff for the new dir approach.
    - switch to an ascii format, maybe sexpr?
    - authdirservers publish blobs of them.
    - other authdirservers fetch these blobs.
    - hidserv people have the option of not uploading their blobs.
    - you can insert a blob via the controller.
    - and there's some amount of backwards compatibility.
    - teach clients, intro points, and hidservs about auth mechanisms.
    - come up with a few more auth mechanisms.
  - auth mechanisms to let hidden service midpoint and responder filter
    connection requests.
  - Bind to random port when making outgoing connections to Tor servers,
    to reduce remote sniping attacks.
  - Have new people be in limbo and need to demonstrate usefulness
    before we approve them.
  - Clients should estimate their skew as median of skew from servers
    over last N seconds.
  - Make router_is_general_exit() a bit smarter once we're sure what it's for.
  - Audit everything to make sure rend and intro points are just as likely to
    be us as not.
  - Do something to prevent spurious EXTEND cells from making middleman
    nodes connect all over.  Rate-limit failed connections, perhaps?
  - Automatically determine what ports are reachable and start using
    those, if circuits aren't working and it's a pattern we recognize
    ("port 443 worked once and port 9001 keeps not working").
  - Limit to 2 dir, 2 OR, N SOCKS connections per IP.
  - Handle full buffers without totally borking
  - Rate-limit OR and directory connections overall and per-IP and
    maybe per subnet.
  - Hold-open-until-flushed now works by accident; it should work by
    design.
  - DoS protection: TLS puzzles, public key ops, bandwidth exhaustion.
    - Specify?
  - tor-resolve script should use socks5 to get better error messages.
  - hidserv offerers shouldn't need to define a SocksPort
    * figure out what breaks for this, and do it.
  - tor should be able to have a pool of outgoing IP addresses
    that it is able to rotate through. (maybe)
    - Specify; implement.
  - let each hidden service (or other thing) specify its own
    OutboundBindAddress?
  - Stop using tor_socketpair to make connection bridges: do an
    implementation that uses buffers only.

Blue-sky:
  - Patch privoxy and socks protocol to pass strings to the browser.
  - Standby/hotswap/redundant hidden services.
  - Robust decentralized storage for hidden service descriptors.
  - The "China problem"
  - Allow small cells and large cells on the same network?
  - Cell buffering and resending. This will allow us to handle broken
    circuits as long as the endpoints don't break, plus will allow
    connection (tls session key) rotation.
  - Implement Morphmix, so we can compare its behavior, complexity, etc.
  - Other transport. HTTP, udp, rdp, airhook, etc. May have to do our own
    link crypto, unless we can bully openssl into it.
  - Need a relay teardown cell, separate from one-way ends.
    (Pending a user who needs this)
  - Handle half-open connections: right now we don't support all TCP
    streams, at least according to the protocol. But we handle all that
    we've seen in the wild.
    (Pending a user who needs this)

Non-Coding:
  - Mark up spec; note unclear points about servers
  - Mention controller libs someplace.
  . more pictures from ren. he wants to describe the tor handshake
NR- write a spec appendix for 'being nice with tor'
  - tor-in-the-media page
  - Remove need for HACKING file.
  - Figure out licenses for website material.
  - Specify the keys and key rotation schedules and stuff

Website:
  - and remove home and make the "Tor" picture be the link to home.
  - put the logo on the website, in source form, so people can put it on
    stickers directly, etc.
R - make a page with the hidden service diagrams.

  - ask Jan to be the translation coordinator? add to volunteer page.

  - add a page for localizing all tor's components.

  - It would be neat if we had a single place that described _all_ the
    tor-related tools you can use, and what they give you, and how well they
    work.  Right now, we don't give a lot of guidance wrt
    torbutton/foxproxy/privoxy/polipo in any consistent place.