On Fri, Mar 20, 2015 at 11:55 AM, Nick Mathewson nickm@torproject.org wrote:
Hi! I've got an end-of-month deliverable to flesh out as many good ideas here as I can, and I'd appreciate feedback on what kind of features it would be good to add to the controller protocol in order to better support testing.
More ideas would be most welcome.
Thanks for the comments, all!
Here's the latest version, with more things folded in, and some syntax sketches.
All syntaxes below are examples meant to focus discussion around what might appear in a spec, not final (ha!) formal specification.
IDEAS =====
1. Step-by-step hidden service connections
Add the ability to create connections to hidden services step by step, to best
What's necessary here is commands to: * Establish a rendezvous point on a given circuit. * Construct and send an introduce1 cell on a given circuit. * Realize that a rendezvous circuit has been constructed.
ESTABLISH_REND {circuit_num} {rend_token}
SEND_INTRO1 {circuit_num} intro_key={key} authinfo={ nfo}
650 RECEIVED_REND_OK {circuit_num}
2. Send a single cell on a circuit
(TESTING ONLY)
For fuzzing and low-level testing purposes, it would be handy to be able to send a single cell on a tor circuit.
This might be better to expose via a low-level modular API than via the control port.
SEND_RAW_RELAY_CELL {circuit_num} LAYER={cpath layer} DATA={hex}
3. Intercept cell by cell on a circuit
(TESTING ONLY)
For fuzzing, testing, and debugging purposes, it might be handy for a controller to be able to observe data cell by cell on a circuit of interest.
This might be better to expose via a low-level modular API than via the control port.
650 RELAY_CELL {circuit_num} LAYER={cpath layer} DATA={hex}
4. Send a single cell on a connection.
(TESTING ONLY)
As 2, but for connections. Note that we might even, for testing, expose this at a sub-cell level.
SEND_RAW_CELL {conn id} DATA={hex}
5. Intercept all cells on a connection
(TESTING ONLY)
As 2, but for connections.
650 CONN_CELL {conn id} DATA={hex}
6. Plug-in to handle a relay or other command.
Right now, all Tor's features need to be baked into Tor; it's not easy to write extensions. We could change that by having the controller able to intersect particular relay or extension commands and act accordingly. This could be used for prototyping new features, etc.
HANDLE_EXTENSION {conn=id|circ=id|conn=*|circ=*} {command_num}
650 EXTENSION {conn=id|circ=id} {command_num} DATA
7. Force a given protocol on a given connection
We could add a feature to restrict what protocols can be negotiated on a given connection we create. This could help us better test our protocols for interoperatbility.
OPEN_CONN {addr=address...} {protos=1,2,3...} {KEYID1=...}
8. Examine fine-grained connection detail.
There are many data available for a given connection (such as fine-grained TLS information) that are not currently exposed on the GETINFO interface. We could make most of this available for testing, pending security analysis.
GETINFO orconn/<id>/all GETINFO orconn/<id>/item
9. Examine cache in detail
In the past we've seen crazy issues with our descriptor caching code. It might be good to expose for testing information about where exactly descriptors are stored, what attributes are set on them, and so on. We could also expose events for cache compaction and discarded expired descriptors.
This might also (TESTING ONLY) extend to events for access of cache documents, information about HS Descs, etc, though that's a bit worrisome.
GETINFO cache/<itemtype>/keys GETINFO cache/<itemtype>/by-key/<key>/ ... raw annotations location
650 CACHE_COMPACTED (bytes_saved) 650 CACHE_DISCARDED (item)
10. Fetch literal documents
Currently there's no way for a controller to ask Tor to download a given descriptor or microdescriptor or networkstatus. That could change.
DOWNLOAD <itemtype> <keys> [{circ=circ_id} {anon=0|1} {cache=router_id}]
11. OOM stats
To resist out-of-memory attacks, we track our memory usage and kill off circuits as needed when memory gets low. We could expose the memory thresholds and current sizes via one or more controller commands.
GETINFO mem-usage/all GETINFO mem-usage/<type>
650 MEM_USAGE k=v k=v k=v 650 MEM_FREED k=v k=v k=v
12. Timeout values
Tor has a truly huge variety of internal timers to ensure that given periodic events happen enough; we could expose those, and (TESTING ONLY) allow controllers to adjust them or trigger their corresponding events.
GETINFO timer/list GETINFO timer/expires/<timer-name>
ADJTIMER <timer-name> <seconds>.<msec>
650 TIMER <timer-name> next=<seconds.msec>
13. Detailed connection debugging info
Current connection events expose only large-scale state changes in connections; we could instead expose every state transition at the cell-by-cell handshake level.
14. Detailed circuit debugging info
As 13 but for circuits.
15. Halt main loop except for control layer.
(TESTING ONLY)
For inspection/debugging purposes, it might be clever to have Tor be able to freeze itself, except for the control layer, and let the controller inquire about information.
This presents implementation challenges, and is probably not a great idea to do before a _big_ refactoring.
FREEZE {LOOP|CONN=id|CIRCUIT=id} UNFREEZE {LOOP|CONN=id|CIRCUIT=id}
16. Service a single connection
(TESTING ONLY)
Currently controllers can disable circuit construction or stream attachment, and do them manually. We might also do this for connections, allowing a testing controller to trace what Tor does cell by cell on a single connection.
STEP {LOOP|CONN=id|CIRCUIT=id}
17. All rephist data
There are many data about history and usage in rephist.c (which stands for 'reputation and history!'). We could expose them, to let us better test them. Some of this might be useful for Seth (previously arm) users.
Spec: This would use GETINFO extensions, and probably some new events.
18. All ratelim data
Sometimes our rate-limiting code can get wonky. It would be great to expose it to Tor controllers in order to help ensure it's behaving correctly. This would include send/receive windows and bw stuff.
Spec: This would use GETINFO extensions, and probably some new events.
19. All accounting data
As 17, but for hibernate.c, which performs bandwidth accounting.
20. All guard transitions
Our guard node state logic is very complicated, and much in need of testing and refactoring. Exposing more state transitions and guard selection transitions to the controller might help. (We have a "GUARD" event now, but it is a bit out of sync with the main implementation)
21. All key transitions
We could generate events every time we change keys, and (TESTING ONLY) allow a controller to time-out a key early.
22. Examine mux settings
(TESTING ONLY)
The circuitmux code that we use to decide which cell to send next is very complex; it would be good to expose its thinking and decisionmaking to a testing controller for better observation.
23. Examine pathbias settings
Our pathbias code is also complex, and a bit flakier than the circuitmux code. We could do for it as with 22 above.
24. Examine cpuworker queues
As of 0.2.6, we have a new cpuworker infrastructure that better sends data to worker threads, but not much visibility into how well it's working. Exposing some information about this to the controller could help us tune better.
650 QUEUED_WORK TYPE=type ID=id RAW=hex NUM_QUEUED=n 650 FINISHED_WORK TYPE=type ID=id STATUS=status RAW=raw
GETINFO cpuworker/all/n_queued GETINFO cpuworker/all/rate GETINFO cpuworker/<num>/... GETINFO cpuworker/<type>/...
25. All geoip data
As 17, but for any geoip information we aren't currently exporting.
26. Replay detection
For hidden service security (and maybe eventually for secure ntor Y reuse) we have to keep replay caches to prevent us from being tricked into handling the same value twice. We don't expose the load on these caches to the controller, however. And we could, to help us better tune them into using a good memory/error-rate trade-off.
GETINFO replaycache/<type>/load GETINFO replaycache/<type>/size
27. Hidden service intropoint changes, desc changes, uploads (+2)
Many hidden service transitions currently generate no events. We could at minimum generate events for changed inroduction points, changed hidden service descriptors, uploading our own HS descriptor.
650 HSSERV <id> <eventtype> k=v k=v k=v
28. Descriptor uploads.
We have an event for when our descriptor has changed, but not for a successful upload for it. We could fix that.
650 UPLOADED <type> <authority> <status>
29. Path generation logic -- expose, allow.
Currently a controller's only visibility into path selection logic is in its outputs, and in the opportunity to replace path selection logic entirely. We could expose more details about the algorithm's operation in order to help better test our path selection.
GEN_PATH <k=v> <k=v>
650 WANT_PATH id <k=v> USE_PATH id node,node,node
30. All PT status information.
Pluggable transport feedback is, at present, very coarse-grained. For testing we might expose more.
31. Crypto operation counts.
We ought to keep count of our various cryptographic operations, and expose them to the controller. This would help us know where to spend our optimization efforts.
GETINFO crypto/{<op>|all}/{<use>|all}/count GETINFO crypto/{<op>|all}/{<use>|all}/speed
32. Forget cached information (+1)
To better test our download logic, it would be helpful to have a way to drop items from our caches.
DROPCACHE {cachetype}
33. Expose number of bytes sent-received per circuit
(TESTING ONLY)
This could help us improve efficiency by investigating slow-circuit issues.
650 CIRCBYTES sent={n} rcvd={n}
34. Modify state of existing connections and circuits
(TESTING ONLY)
This could help with fault injection to better validate error-handling code.
IMPLEMENTATION NOTES: =====================
Not lightly does one list 33 controller improvement areas. If we're hoping to do these without too much programmer time and , we need to take a much more principled approach to implementing controller commands. For #8351 I worked on a branch called 'ticket8351' that has some code we could use here.
COMPATIBILITY NOTES: ====================
Many of the features here are ones that we might not want to promise to support indefinitely; we should gate them behind a USEFEATURE command, and maybe place them in an annex of the control spec.