tor/doc/TODO
Roger Dingledine ff76072feb more todo action, and a spec note
svn:r1516
2004-04-07 06:40:06 +00:00

386 lines
19 KiB
Plaintext

Legend:
SPEC!! - Not specified
SPEC - Spec not finalized
NICK - nick claims
ARMA - arma claims
- Not done
* Top priority
. Partially done
o Done
D Deferred
X Abandoned
Bugs:
o we call signal(), but we should be calling sigaction()
o send socks rejects when things go bad ?
o on solaris, need to build with
LDFLAGS="-lsocket -lnsl" ./configure
o on solaris, we HAVE_UNAME but the uname() call fails?
. should maybe make clients exit(1) when bad things happen?
e.g. clock skew.
o client-side dns cache doesn't appear to be getting populated
by 'connected' cells. In fact, the 'connected' cells don't even
include the IP.
o When it can't resolve any dirservers, it is useless from then on.
We should make it reload the RouterFile if it has no dirservers.
o Sometimes it picks a middleman node as the exit for a circuit.
o if you specify a non-dirserver as exitnode or entrynode, when it
makes the first few circuits it hasn't yet fetched the directory,
so it warns that it doesn't know the node.
o make 'make test' exit(1) if a test fails.
. fix buffer unit test so it passes
Short-term:
- add in 'notice' log level
- make recommendedversions different for clients and servers.
e.g. C0.0.3 vs S0.0.3?
o put IP into the descriptor, so clients don't need to resolve things
o when you hup, rewrite the router.desc file (and maybe others)
- consider handling broken socks4 implementations
- improve how it behaves when i remove a line from the approved-routers files
- Make tls connections tls_close intentionally
- Add '[...truncated]' or similar to truncated log entries (like the directory
in connection_dir_process_inbuf()).
. Make logs handle it better when writing to them fails.
o leave server descriptor out of directory if it's too old
o Rename ACI to circID
o integrate rep_ok functions, see what breaks
- update tor faq
o obey SocksBindAddress, ORBindAddress
o warn if we're running as root
o make connection_flush_buf() more obviously obsolete
o let hup reread the config file, eg so we can get new exit
policies without restarting
o Put recommended_versions in a config entry
X use times(2) rather than gettimeofday to measure how long it
takes to process a cell
o Separate trying to rebuild a circuit because you have none from trying
to rebuild a circuit because the current one is stale
X Continue reading from socks port even while waiting for connect.
o Exit policies
o Spec how to write the exit policies
o Path selection algorithms
o Choose path more incrementally
o Let user request first/last node
o And disallow certain nodes
D Choose path by jurisdiction, etc?
o Make relay end cells have failure status and payload attached
X let non-approved routers handshake.
X Dirserver shouldn't put you in running-routers list if you haven't
uploaded a descriptor recently
X migrate to using nickname rather than addr:port for routers
- migrate to using IPv6 sizes everywhere
o Move from onions to ephemeral DH
o incremental path building
o transition circuit-level sendmes to hop-level sendmes
o implement truncate, truncated
o move from 192byte DH to 128byte DH, so it isn't so damn slow
X exiting from not-last hop
X OP logic to decide to extend/truncate a path
X make sure exiting from the not-last hop works
X logic to find last *open* hop, not last hop, in cpath
o Remember address and port when beginning.
- Extend by nickname/hostname/something, not by IP.
- Need a relay teardown cell, separate from one-way ends.
X remove per-connection rate limiting
- Make it harder to circumvent bandwidth caps: look at number of bytes
sent across sockets, not number sent inside TLS stream.
o Audit users of connnection_remove and connection_free outside of
main.c; many should use mark_for_close instead.
Rendezvous service:
o Design and specify protocol
o Possible preliminary refactoring:
o Should we break circuits up into "circuit-with-cpath" and
"circuit-without-cpath"?
o We need a way to tag circuits as special-purpose circuits for:
o Connecting from Bob's OP to the introduction point
o Sending introduction requests from the IPoint to Bob
o Connecting from Alice to the rendezvous point for Bob
o Connecting from Bob to the rendezvous point for Alice
o Waiting at a rendezvous point to be joined
o Joined to another circuit at the rendezvous point.
(We should also enumerate all the states that these operations
can be in.) [NM]
o Add circuit metadata [NM]
o Code to configure hidden services [NM] 4 hours
o Service descriptors
o OPs need to maintain identity keys for hidden services [NM]
o Code to generate and parse service descriptors [NM]
o Advertisement
o Generate y.onion hostnames [NM]
o Store y.onion hostnames to disk. [NM]
o Code to do an HTTP connection over Tor from within Tor [RD]
o Publish service descriptors to directory [RD]
o Directory accepts and remembers service descriptors, and
delivers them as requested
o Frontend [RD]
o Backend [NM]
o Code for OPs to retrieve (and cache?) service descriptors [RD]
o Rendezvous
o Code as needed to generate and parse all rendezvous-related
cell types, and do all handshaking [NM]
o ORs implement introduction points
o OPs with hidden services establish introduction points
o ORs implement rendezvous points
o OPs notice y.onion URLs, and:
o Retrieve service descriptors
o Establish rendezvous points
o Send introduction requests to introduction points
o Communication
o OPs remember which circuits are used for which rendezvous
points, and can look up circuits by location-hidden service
o OPs send/handle BEGIN cells for location-hidden services
o End-to-end communication for location-hidden services
- a section in the man pages: how to configure hidden services
o let bob use himself as a rendezvous point
- let bob choose himself as intro point
- let bob replenish his intro points and republish
- alice retries introduction and rendezvous a few times?
- should alice ever try to refresh her service desc cache entries?
should she expire them after e.g. 15 mins?
- race condition: alice has the serverdesc in her cache, she opens
the circs, serverdesc expires and is flushed, then she goes
to send the intro cell. should serverdesc cache have a
last-touched field? are there better fixes?
- backward compatibility: when only certain nodes know about rend
protocol, how do we deal? have nodes parse the tor version field?
force an upgrade? simply be more robust against useless nodes?
On-going
. Better comments for functions!
. Go through log messages, reduce confusing error messages.
. make the logs include more info (fd, etc)
. Unit tests
. Update the spec so it matches the code
Mid-term:
- Refactor: add own routerinfo to routerlist. Right now, only
router_get_by_nickname knows about 'this router', as a hack to
get circuit_launch_new to do the right thing.
- Rotate tls-level connections -- make new ones, expire old ones.
So we get actual key rotation, not just symmetric key rotation
o Are there anonymity issues with sequential streamIDs? Sequential
circIDs? Eg an attacker can learn how many there have been.
The fix is to initialize them randomly rather than at 1.
- Look at having smallcells and largecells
. Redo scheduler
o fix SSL_read bug for buffered records
- make round-robining more fair
- What happens when a circuit's length is 1? What breaks?
. streams / circuits
o Implement streams
o Rotate circuits after N minutes?
X Circuits should expire when circuit->expire triggers
NICK . Handle half-open connections
o openssh is an application that uses half-open connections
o Figure out what causes connections to close, standardize
when we mark a connection vs when we tear it down
o Look at what ssl does to keep from mutating data streams
o Put CPU workers in separate processes
o Handle multiple cpu workers (one for each cpu, plus one)
o Queue for pending tasks if all workers full
o Support the 'process this onion' task
D Merge dnsworkers and cpuworkers to some extent
o Handle cpuworkers dying
. Scrubbing proxies
- Find an smtp proxy?
- Check the old smtp proxy code
o Find an ftp proxy? wget --passive
D Wait until there are packet redirectors for Linux
. Get socks4a support into Mozilla
. Tests
o Testing harness/infrastructure
D System tests (how?)
- Performance tests, so we know when we've improved
. webload infrastructure (Bruce)
. httperf infrastructure (easy to set up)
. oprofile (installed in RH >8.0)
NICK . Daemonize and package
o Teach it to fork and background
. Red Hat spec file
o Debian spec file equivalent
. Portability
. Which .h files are we actually using?
. Port to:
o Linux
o BSD
. Solaris
o Cygwin
. Win32
o OS X
- deal with pollhup / reached_eof on all platforms
o openssl randomness
o inet_ntoa
o stdint.h
- Make a script to set up a local network on your machine
o More flexibility in node addressing
D Support IPv6 rather than just 4
o Handle multihomed servers (config variable to set IP)
In the distant future:
D Load balancing between router twins
D Keep track of load over links/nodes, to
know who's hosed
SPEC!! D Non-clique topologies
D Implement our own memory management, at least for common structs
(Not ever necessary?)
D Advanced directory servers
D Automated reputation management
SPEC!! D Figure out how to do threshold directory servers
D jurisdiction info in dirserver entries? other info?
Older (done) todo stuff:
For 0.0.2pre17:
o Put a H(K | handshake) into the onionskin response
o Make cells 512 bytes
o Reduce streamid footprint from 7 bytes to 2 bytes
X Check for collisions in streamid (now possible with
just 2 bytes), and back up & replace with padding if so
o Use the 4 reserved bytes in each cell header to keep 1/5
of a sha1 of the ongoing relay payload (move into stream header)
o Move length into the stream header too
o Make length 2 bytes
D increase DH key length
D increase RSA key length
D Spec the stream_id stuff. Clarify that nobody on the backward
stream should look at stream_id.
Cell:
ACI (anonymous circuit identifier) [2 bytes]
Command [1 byte]
Payload (padded with 0 bytes) [509 bytes]
Relay payload:
Relay command [1 byte]
Stream ID [7 bytes]
Partial SHA-1 [4 bytes]
Length [2 bytes]
Relay payload [495 bytes]
For 0.0.2pre15:
o don't pick exit nodes which will certainly reject all things.
o don't pick nodes that the directory says are down
o choose randomly from running dirservers, not just first one
o install the man page
o warn when client-side tries an address/port which no router in the dir accepts.
For 0.0.2pre14:
o More flexible exit policies (18.*, 18.0.0.0/8)
o Work to succeed in the precense of exit policy violation
o Replace desired_path_len with opaque path-selection specifier
o Client-side DNS caching
o Add entries to client DNS cache based on END cells
o Remove port from END_REASON_EXITPOLICY cells
o Start building new circuits when we get an exit-policy
failure. (Defer exiting from the middle of existing
circuits or extending existing circuits for later.)
o Implement function to check whether a routerinfo_t
supports a given exit addr.
o Choose the exit node of an in-progress circuit based on
pending AP connections.
o Choose the exit node _first_, then beginning, then
middle nodes.
Previous:
o Get tor to act like a socks server
o socks4, socks4a
o socks5
o routers have identity key, link key, onion key.
o link key certs are
D signed by identity key
D not in descriptor
o not in config
D not on disk
o identity and onion keys are in descriptor (and disk)
o upon boot, if it doesn't find identity key, generate it and write it.
o also write a file with the identity key fingerprint in it
o router generates descriptor: flesh out router_get_my_descriptor()
o Routers sign descriptors with identity key
o routers put version number in descriptor
o routers should maybe have `uname -a` in descriptor?
o Give nicknames to routers
o in config
o in descriptors
o router posts descriptor
o when it boots
o every DirFetchPostPeriod seconds
D when it changes
o change tls stuff so certs don't get written to disk, or read from disk
o make directory.c 'thread'safe
o dirserver parses descriptor
o dirserver checks signature
D client checks signature?
o dirserver writes directory to file
o reads that file upon boot
o directory includes all routers, up and down
o add "up" line to directory, listing nicknames
o instruments ORs to report stats
o average cell fullness
o average bandwidth used
o configure log files. separate log file, separate severities.
o what assumptions break if we fclose(0) when we daemonize?
o make buffer struct elements opaque outside buffers.c
o add log convention to the HACKING file
o make 'make install' do the right thing
o change binary name to tor
o change config files so you look at commandline, else look in
/etc/torrc. no cascading.
o have an absolute datadir with fixed names for files, and fixed-name
keydir under that with fixed names
o Move (most of) the router/directory code out of main.c
o Simple directory servers
o Include key in source; sign directories
o Signed directory backend
o Document
o Integrate
o Add versions to code
o Have directories list recommended-versions
o Include line in directories
o Check for presence of line.
o Quit if running the wrong version
o Command-line option to override quit
o Add more information to directory server entries
o Exit policies
o Clearer bandwidth management
o Do we want to remove bandwidth from OR handshakes?
o What about OP handshakes?
X Move away from openssl
o Abstract out crypto calls
X Look at nss, others? Just include code?
o Use a stronger cipher
o aes now, by including the code ourselves
X On the fly compression of each stream
o Clean up the event loop (optimize and sanitize)
o Remove that awful concept of 'roles'
o Terminology
o Circuits, topics, cells stay named that
o 'Connection' gets divided, or renamed, or something?
o DNS farm
o Distribute queries onto the farm, get answers
o Preemptively grow a new worker before he's needed
o Prune workers when too many are idle
o DNS cache
o Clear DNS cache over time
D Honor DNS TTL info (how??)
o Have strategy when all workers are busy
o Keep track of which connections are in dns_wait
o Need to cache positives/negatives on the tor side
o Keep track of which queries have been asked
o Better error handling when
o An address doesn't resolve
o We have max workers running
o Consider taking the master out of the loop?
X Implement reply onions
o Total rate limiting
o Look at OR handshake in more detail
o Spec it
o Merge OR and OP handshakes
o rearrange connection_or so it doesn't suck so much to read
D Periodic link key rotation. Spec?
o wrap malloc with something that explodes when it fails
o Clean up the number of places that get to look at prkey