[This is a significantly revised version of the last version of this proposal draft, sent here for comment.
The last version of this draft was https://lists.torproject.org/pipermail/tor-dev/2015-September/009587.html . This is more or less a complete rewrite, after I convinced myself that the original ideas wouldn't work.
Part of the functionality of that old draft has been subsumed in proposal 264.]
1. Introduction
Frequently, we find that very old versions of Tor should no longer be supported on the network. To remove relays is easy enough: we simply update the directory authorities to stop listing relays that advertise versions that are too old.
But to disable clients is harder.
In another proposal I describe a system for letting future clients go gracefully obsolete. This proposal explains how we can safely disable the obsolete clients we have today (and all other client versions of Tor to date, assuming that they will someday become obsolete).
1.1. Why disable clients?
* Security. Anybody who hasn't updated their Tor client in 5 years is probably vulnerable to who-knows-what attacks. They aren't likely to get much anonymity either.
* Withstand zombie installations. Some Tors out there were once configured to start-on-boot systems that are now unmaintained. (See 1.4 below.) They put needless load on the network, and help nobody.
* Be able to remove backward-compatibility code. Currently, Tor supports some truly ancient protocols in order to avoid breaking ancient versions or Tor. This code needs to be maintained and tested. Some of it depends on undocumented or deprecated or non-portable OpenSSL features, and makes it hard to produce a conforming Tor server implementation.
* Make it easier to write a conforming Tor relay. If a Tor relay needs to support every Tor client back through the beginning of time, that makes it harder to develop and test compatible implementations.
1.2. Is this dangerous?
I don't think so. This proposal describes a way to make older clients gracefully disconnect from the network only when a majority of authorities agree that they should. A majority of authorities already have the ability to inflict arbitrary degrees of sabotage on the consensus document.
1.3. History
The earliest versions of Tor checked the recommended-versions field in the directory to see whether they should keep running. If they saw that their version wasn't recommended, they'd shut down. There was an "IgnoreVersion" option that let you keep running anyway.
Later, around 2004, the rule changed to "shut down if the version is _obsolete_", where obsolete was defined as "not recommended, and older than a version that is recommended."
In 0.1.1.7-alpha, we made obsolete versions only produce a warning, and removed IgnoreVersion. (See 3ac34ae3293ceb0f2b8c49.)
We have still disabled old tor versions. With Tor 0.2.0.5-alpha, we disabled Tor versions before 0.1.1.6-alpha by having the v1 authorities begin publishing empty directories only.
In version 0.2.5.2-alpha, we completely removed support for the v2 directory protocol used before Tor 0.2.0; there are no longer any v2 authorities on the network.
Tor versions before 0.2.1 will currently not progress past fetching an initial directory, because they believe in a number of directory authority identity keys that no longer sign the directory.
Tor versions before 0.2.4 are (lightly) throttled in multihop circuit creation, because we prioritize ntor CREATE cells over TAP ones when under load.
1.4. The big problem: slow zombies and fast zombies
It would be easy enough to 'disable' old clients by simply removing server support for the obsolete protocols that they use. But there's a problem with that approach: what will the clients do when they fail to make connections, or to extend circuits, or whatever else they are no longer able to do?
* Ideally, I'd like such clients to stop functioning _quietly_. If they stop contacting the network, that would be best.
* Next best would be if these clients contacted the network only occasionally and at different times. I'll call these clients "slow zombies".
* Worse would be if the clients contact the network frequently, over and over. I'll call these clients "fast zombies". They would be at their worst when they focus on authorities, or when they act in synchrony to all strike at once.
One goal of this proposal is to ensure that future clients to not become zombies at all; and that ancient clients become slow zombies at worst.
2. Some ideas that don't work.
2.1. Dropping connections based on link protocols.
Tor versions before before 0.2.3.6-alpha use a renegotiation-based handshake instead of our current handshake. We could detect these handshakes and close the connection at the relay side if the client attempts to renegotiate.
I've tested these changes on versions maint-0.2.0 through maint-0.2.2. They result in zombies with the following behavior:
The client contact each authority it knows about, attempting to make a one-hop directory connection. It fails, detects a failure, then reconnects more and more slowly ... but one hour later, it resets its connection schedule and starts again.
In the steady state this appears to result in about two connections per client per authority per hour. That is probably too many.
(Most authorities would be affected: of the authorities that existed in 0.2.2, gabelmoo has moved and turtles has shut down. The authorities Faravahar and longclaw are new. The authorities moria1, tor26, dizum, dannenberg, urras, maatuska and maatuska would all get hit here.)
(We could simply remove the renegotiation-detection code entirely, and reply to all connections with an immediate VERSIONS cell. The behavior would probably be the same, though.)
If we throttled connections rather than closing them, we'd only get one connnection per authority per hour, but authorities would have to keep open a potentially huge number of sockets.
2.2. Blocking circuit creation under certain circumstances
In tor 0.2.5.1-alpha, we began ignoring the UseNTorHandshake option, and always preferring the ntor handshake where available.
Unfortunately, we can't simply drop all TAP handshakes, since clients and relays can still use them in the hidden service protocol. But we could detect these versions by:
Looking for use of a TAP handshake from an IP not associated with with any known relay, or on a connection where the client did not authenticate. (This could be from a bridge, but clients don't build circuits that go to an IntroPoint or RendPoint directly after a bridge.)
This would still result in clients not having directories, however, and retrying once an hours.
3. Ideas that might work
3.1. Move all authorities to new ports
We could have each authority known to older clients start listening for connections at a new port P. We'd forward the old port to the new port. Once sufficiently many clients were using the new ports, we could disable the forwarding.
This would result in the old clients turning into zombies as above, but they would only be scrabbling at nonexistent ports, causing less load on the authorities.
[This proposal would probably be easiest to implement.]
3.2. Start disabling old link protocols on relays
We could have new relays start dropping support for the old link protocols, while maintaining support on the authorities and older relays.
The result here would be a degradation of older client performance over time. They'd still behave zombieishly if the authorities dropped support, however.
3.3. Changing the consensus format.
We could allow 'f' (short for "flag") as a synonym for 's' in consensus documents. Later, if we want to disable all Tor versions before today, we can change the consensus algorithm so that the consensus (or perhaps only the microdesc consensus) is spelled with 'f' lines instead of 'f' lines. This will create a consensus which older clients and relays parse as having all nodes down, which will make them not connect to the network at all.
We could similarly replace "r" with "n", or replace Running with Online, or so on.
In doing this, we could also rename fresh-until and valid-until, so that new clients would have the real expiration date, and old clients would see "this consensus never expires". This would prevent them from downloading new consensuses.
[This proposal would result in the quietest shutdown.]