$Id$ Tor Path Specification Roger Dingledine Nick Mathewson Note: This is an attempt to specify Tor as currently implemented. Future versions of Tor will implement improved algorithms. This document tries to cover how Tor chooses to build circuits and assign streams to circuits. Other implementations MAY take other approaches, but implementors should be aware of the anonymity and load-balancing implications of their choices. THIS SPEC ISN'T DONE OR CORRECT YET. 1. General operation Tor begins building circuits as soon as it has enough directory information to do so (see section 5.1 of dir-spec.txt). Some circuits are built preemptively because we expect to need them later (for user traffic), and some are built because of immediate need (for user traffic that no current circuit can handle, for testing the network or our reachability, and so on). When a client application creates a new stream (by opening a SOCKS connection or launching a resolve request), we attach it to an appropriate open circuit if one exists, or wait if an appropriate circuit is in-progress. We launch a new circuit only if no current circuit can handle the request. We rotate circuits over time to avoid some profiling attacks. To build a circuit, we choose all the nodes we want to use, and then construct the circuit. Sometimes, when we want a circuit that ends at a given hop, and we have an appropriate unused circuit, we "cannibalize" the existing circuit and extend it to the new terminus. These processes are described in more detail below. This document describes Tor's automatic path selection logic only; path selection can be overridden by a controller (with the EXTENDCIRCUIT and ATTACHSTREAM commands). Paths constructed through these means may violate some constraints given below. 1b. Terminology A "path" is an ordered sequence of nodes, not yet built as a circuit. A "clean" circuit is one that has not yet been used for any traffic. A "fast" or "stable" or "valid" node is one that has the 'Fast' or 'Stable' or 'Valid' flag set respectively, based on our current directory information. A "fast" or "stable" circuit is one consisting only of "fast" or "stable" nodes. In an "exit" circuit, the final node is chosen based on waiting stream requests if any, and in any case it avoids nodes with exit policy of "reject *:*". An "internal" circuit, on the other hand, is one where the final node is chosen just like a middle node (ignoring its exit policy). A "request" is a client-side stream or DNS resolve that needs to be served by a circuit. A "pending" circuit is one that we have started to build, but which has not yet completed. A circuit or path "supports" a request if it is okay to use the circuit/path to fulfill the request, according to the rules given below. A circuit or path "might support" a request if some aspect of the request is unknown (usually its target IP), but we believe the path probably supports the request according to the rules given below. 2. Building circuits 2.1. When we build. 2.1.1. Clients build circuits preemptively When running as a client, Tor tries to maintain at least a certain number of clean circuits, so that new streams can be handled quickly. To increase the likelihood of success, Tor tries to predict what circuits will be useful by choosing from among nodes that support the ports we have used in the recent past (by default one hour). Specifically, on startup Tor tries to maintain one clean fast exit circuit that allows connections to port 80, and at least two internal circuits in case we get a resolve request or hidden service request (at least three internal circuits if we _run_ a hidden service). After that, Tor will adapt the circuits that it preemptively builds based on the requests it sees from the user: it tries to have a clean fast exit circuit available for every port seen recently (one circuit is adequate for many predicted ports -- it doesn't keep a separate circuit for each port), and it tries to have the above internal circuits available if we've seen resolves or hidden service activity recently. If there are 12 clean circuits open, it doesn't open more even if it has more predictions. Lastly, note that if there are no requests from the user for an hour, Tor will predict no use and build no preemptive circuits. The Tor client SHOULD NOT store its list of predicted requests to a persistent medium. 2.1.2. Clients build circuits on demand Additionally, when a client request exists that no circuit (built or pending) might support, we create a new circuit to support the request. We do so by picking a request arbitrarily, launching a circuit to support it, and repeating until every unattached request might be supported by a pending or built circuit. For hidden service interations, we can "cannibalize" a clean internal circuit if one is available, so we don't need to build those circuits from scratch on demand. We can also cannibalize clean circuits when the client asks to exit at a given node -- either via mapaddress or the ".exit" notation, or because the destination is running at the same location as an exit node. 2.1.3. Servers build circuits for testing reachability Tor servers test reachability of their ORPort on start and whenever their IP address changes. XXXX 2.1.4. Hidden-service circuits See section 4 below. 2.1.5. Rate limiting of failed circuits If we fail to build a circuit N times in a X second period (see Section 2.3 for how this works), we stop building circuits until the X seconds have elapsed. XXX 2.1.6. When to tear down circuits 2.2. Path selection and constraints We choose the path for each new circuit before we build it. We choose the exit node first, followed by the other nodes in the circuit. All paths we generate obey the following constraints: - We do not choose the same router twice for the same path. - We do not choose any router in the same family as another in the same path. - We do not choose any router in the same /16 subnet as another in the same path (unless EnforceDistinctSubnets is 0). - We don't choose any non-running or non-valid router unless we have been configured to do so. By default, we are configured to allow non-valid routers in "middle" and "rendezvous" positions. - If we're using Guard nodes, the first node must be a Guard (see 5 below) - XXXX Choosing the length For circuits that do not need to be not "fast", when choosing among multiple candidates for a path element, we choose randomly. For "fast" circuits, we pick a given router as an exit with probability proportional to its advertised bandwidth [the smaller of the 'rate' and 'observed' arguments to the "bandwidth" element in its descriptor]. If a router's advertised bandwidth is greater than MAX_BELIEVABLE_BANDWIDTH (1.5 MB/s), we clip to that value. For non-exit positions on "fast" circuits, we pick routers as above, but we weight the clipped advertised bandwidth of Exit-flagged nodes depending on the fraction of bandwidth available from non-Exit nodes. Call the total clipped advertised bandwidth for Exit nodes under consideration E, and the total clipped advertised bandwidth for non-Exit nodes under consideration N. If E..exit, the request is rewritten to a request for , and the request is only supported by the exit whose nickname or fingerprint is . 2.3. Handling failure If an attempt to extend a circuit fails (either because the first create failed or a subsequent extend failed) then the circuit is torn down and is no longer pending. (XXXX really?) Requests that might have been supported by the pending circuit thus become unsupported, and a new circuit needs to be constructed. If a stream "begin" attempt fails with an EXITPOLICY error, we decide that the exit node's exit policy is not correctly advertised, so we treat the exit node as if it were a non-exit until we retrieve a fresh descriptor for it. XXXX 3. Attaching streams to circuits When a circuit that might support a request is built, Tor tries to attach the request's stream to the circuit and sends a BEGIN or RESOLVE relay cell as appropriate. If the request completes unsuccessfully, Tor considers the reason given in the CLOSE relay cell. [XXX yes, and?] After a request has remained unattached for [XXXX interval?], Tor abandons the attempt and signals an error to the client as appropriate (e.g., by closing the SOCKS connection). XXX Timeouts and when Tor auto-retries. * What stream-end-reasons are appropriate for retrying. If no reply to BEGIN/RESOLVE, then the stream will timeout and fail. 4. Hidden-service related circuits XXX Tracking expected hidden service use (client-side and hidserv-side) 5. Guard nodes XXX writeme 6. Testing circuits (From some emails by arma) Right now the code exists to pick helper nodes, store our choices to disk, and use them for our entry nodes. But there are three topics to tackle before I'm comfortable turning them on by default. First, how to handle churn: since Tor nodes are not always up, and sometimes disappear forever, we need a plan for replacing missing helpers in a safe way. Second, we need a way to distinguish "the network is down" from "all my helpers are down", also in a safe way. Lastly, we need to examine the situation where a client picks three crummy helper nodes and is forever doomed to a lousy Tor experience. Here's my plan: How to handle churn. - Keep track of whether you have ever actually established a connection to each helper. Any helper node in your list that you've never used is ok to drop immediately. Also, we don't save that one to disk. - If all our helpers are down, we need more helper nodes: add a new one to the *end*of our list. Only remove dead ones when they have been gone for a very long time (months). - Pick from the first n (by default 3) helper nodes in your list that are up (according to the network-statuses) and reachable (according to your local firewall config). - This means that order matters when writing/reading them to disk. How to deal with network down. - While all helpers are down/unreachable and there are no established or on-the-way testing circuits, launch a testing circuit. (Do this periodically in the same way we try to establish normal circuits when things are working normally.) (Testing circuits are a special type of circuit, that streams won't attach to by accident.) - When a testing circuit succeeds, mark all helpers up and hold the testing circuit open. - If a connection to a helper succeeds, close all testing circuits. Else mark that helper down and try another. - If the last helper is marked down and we already have a testing circuit established, then add the first hop of that testing circuit to the end of our helper node list, close that testing circuit, and go back to square one. (Actually, rather than closing the testing circuit, can we get away with converting it to a normal circuit and beginning to use it immediately?) How to pick non-sucky helpers. - When we're picking a new helper nodes, don't use ones which aren't reachable according to our local ReachableAddresses configuration. (There's an attack here: if I pick my helper nodes in a very restrictive environment, say "ReachableAddresses 18.0.0.0/255.0.0.0:*", then somebody watching me use the network from another location will guess where I first joined the network. But let's ignore it for now.) - Right now we choose new helpers just like we'd choose any entry node: they must be "stable" (claim >1day uptime) and "fast" (advertise >10kB capacity). In 0.1.1.11-alpha, clients let dirservers define "stable" and "fast" however they like, and they just believe them. So the next step is to make them a function of the current network: e.g. line up all the 'up' nodes in order and declare the top three-quarter to be stable, fast, etc, as long as they meet some minimum too. - If that's not sufficient (it won't be), dirservers should introduce a new status flag: in additional to "stable" and "fast", we should also describe certain nodes as "entry", meaning they are suitable to be chosen as a helper. The first difference would be that we'd demand the top half rather than the top three-quarters. Another requirement would be to look at "mean time between returning" to ensure that these nodes spend most of their time available. (Up for two days straight, once a month, is not good enough.) - Lastly, we need a function, given our current set of helpers and a directory of the rest of the network, that decides when our helper set has become "too crummy" and we need to add more. For example, this could be based on currently advertised capacity of each of our helpers, and it would also be based on the user's preferences of speed vs. security. *** Lasse wrote: > I am a bit concerned with performance if we are to have e.g. two out of > three helper nodes down or unreachable. How often should Tor check if > they are back up and running? Right now Tor believes a threshold of directory servers when deciding whether each server is up. When Tor observes a server to be down (connection failed or building the first hop of the circuit failed), it marks it as down and doesn't try it again, until it gets a new network-status from somebody, at which point it takes a new concensus and marks the appropriate servers as up. According to sec 5.1 of dir-spec.txt, the client will try to fetch a new network-status at least every 30 minutes, and more often in certain cases. With the proposed scheme, we'll also mark all our helpers as up shortly after the last one is marked down. > When should there be > added an extra node to the helper node list? This is kind of an > important threshold? I agree, this is an important question. I don't have a good answer yet. Is it terrible, anonymity-wise, to add a new helper every time only one of your helpers is up? Notice that I say add rather than replace -- so you'd only use this fourth helper when one of your main three helpers is down, and if three of your four are down, you'd add a fifth, but only use it when two of the first four are down, etc. In fact, this may be smarter than just picking a random node for your testing circuit, because if your network goes up and down a lot, then eventually you have a chance of using any entry node in the network for your testing circuit. We have a design choice here. Do we only try to use helpers for the connections that will have streams on them (revealing our communication partners), or do we also want to restrict the overall set of nodes that we'll connect to, to discourage people from enumerating all Tor clients? I'm increasingly of the belief that we want to hide our presence too, based on the fact that Steven and George and others keep coming up with attacks that start with "Assuming we know the set of users". If so, then here's a revised "How to deal with network down" section: 1) When a helper is marked down or the helper list shrinks, and as a result the total number of helpers that are either (up and reachable) or (reachable but never connected to) is <= 1, then pick a new helper and add it to the end of the list. [We count nodes that have never been connected to, since otherwise we might keep on adding new nodes before trying any of them. By "reachable" I mean "is allowed by ReachableAddresses".] 2) When you fail to connect to a helper that has never been connected to, you remove him from the list right then (and the above rule might kick in). 3) When you succeed at connecting to a helper that you've never connected to before, mark all reachable helpers earlier in the list as up, and close that circuit. [We close the circuit, since if the other helpers are now up, we prefer to use them for circuits that will reveal communication partners.] This certainly seems simpler. Are there holes that I'm missing? > If running from a laptop you will meet different firewall settings, so > how should Helper Nodes settings keep up with moving from an open > ReachableAddresses to a FascistFirewall setting after the helper nodes > have been selected? I added the word "reachable" to three places in the above list, and I believe that totally solves this question. And as a bonus, it leads to an answer to Nick's attack ("If I pick my helper nodes all on 18.0.0.0:*, then I move, you'll know where I bootstrapped") -- the answer is to pick your original three helper nodes without regard for reachability. Then the above algorithm will add some more that are reachable for you, and if you move somewhere, it's more likely (though not certain) that some of the originals will become useful. Is that smart or just complex? > What happens if(when?) performance of the third node is bad? My above solution solves this a little bit, in that we always try to have two nodes available. But what if they are both up but bad? I'm not sure. As my previous mail said, we need some function, given our list of helpers and the network directory, that will tell us when we're in a bad situation. I can imagine some simple versions of this function -- for example, when both our working helpers are in the bottom half of the nodes, ranked by capacity. But the hard part: what's the remedy when we decide there's something to fix? Do we add a third, and now we have two crummy ones and a new one? Or do we drop one or both of the bad ones? Perhaps we believe the latest claim from the network-status concensus, and we count a helper the dirservers believe is crummy as "not worth trying" (equivalent to "not reachable under our current ReachableAddresses config") -- and then the above algorithm would end up adding good ones, but we'd go back to the originals if they resume being acceptable? That's an appealing design. I wonder if it will cause the typical Tor user to have a helper node list that comprises most of the network, though. I'm ok with this. > Another point you might want to keep in mind, is the possibility to > reuse the code in order to add a second layer helper node (meaning node > number two) to "protect" the first layer (node number one) helper nodes. > These nodes should be tied to each of the first layer nodes. E.g. there > is one helper node list, as described in your mail, for each of the > first layer nodes, following their create/destroy. True. Does that require us to add a fourth hop to our path length, since the first hop is from a limited set, the second hop is from a limited set, and the third hop might also be constrained because, say, we're asking for an unusual exit port? > Another of the things might worth adding to the to do list is > localization of server (helper) nodes. Making it possible to pick > countries/regions where you do (not) want your helper nodes located. (As > in "HelperNodesLocated us,!eu" etc.) I know this requires the use of > external data and may not be worth it, but it _could_ be integrated at > the directory servers only -- adding a list of node IP's and e.g. a > country/region code to the directory and thus reduce the overhead. (?) > Maybe extending the Family-term? I think we are heading towards doing path selection based on geography, but I don't have a good sense yet of how that will actually turn out -- that is, with what mechanism Tor clients will learn the information they need. But this seems to be something that is orthogonal to the rest of this discussion, so I look forward to having somebody else solve it for us, and fitting it in when it's ready. :) > And I would like to keep an option to pick the first X helper nodes > myself and then let Tor extend this list if these nodes are down (like > EntryNodes in current code). Even if this opens up for some new types of > "relationship" attacks. Good idea. Here's how I'd like to name these: The "EntryNodes" config option is a list of seed helper nodes. When we read EntryNodes, any node listed in entrynodes but not in the current helper node list gets *pre*pended to the helper node list. The "NumEntryNodes" config option (currently called NumHelperNodes) specifies the number of up, reachable, good-enough helper nodes that will make up the pool of possible choices for first hop, counted from the front of the helper node list until we have enough. The "UseEntryNodes" config option (currently called UseHelperNodes) tells us to turn on all this helper node behavior. If you set EntryNodes, then this option is implied. The "StrictEntryNodes" config option, provided for backward compatibility and for debugging, means a) we replace the helper node list with the current EntryNodes list, and b) whenever we would do an operation that alters the helper node list, we don't. (Yes, this means that if all the helper nodes are down, we lose until we mark them up again. But this is how it behaves now.) > I am sure my next point has been asked before, but what about testing > the current speed of the connections when looking for new helper nodes, > not only testing the connectivity? I know this might contribute to a lot > of overhead in the network, but if this only occur e.g. when using > helper nodes as a Hidden Service it might not have that large an impact, > but could help availability for the services? If we're just going to be testing them when we're first picking them, then it seems we can do the same thing by letting the directory servers test them. This has the added benefit that all the (behaving) clients use the same data, so they don't end up partitioned by a node that (for example) performs selectively for his victims. Another idea would be to periodically keep track of what speeds you get through your helpers, and make decisions from this. The reason we haven't done this yet is because there are a lot of variables -- perhaps the web site is slow, perhaps some other node in the path is slow, perhaps your local network is slow briefly, perhaps you got unlucky, etc. I believe that over time (assuming the user has roughly the same browsing habits) all of these would average out and you'd get a usable answer, but I don't have a good sense of how long it would take to converge, so I don't know whether this would be worthwhile. > BTW. I feel confortable with all the terms helper/entry/contact nodes, > but I think you (the developers) should just pick one and stay with it > to avoid confusion. I think I'm going to try to co-opt the term 'Entry' node for this purpose. We're going to have to keep referring to helper nodes for the research community for a while though, so they realize that Tor does more than just let users ask for certain entry nodes. ============================================================ Some stuff that worries me about entry guards. 2006 Jun, Nickm. 1. It is unlikely for two users to have the same set of entry guards. 2. Observing a user is sufficient to learn its entry guards. 3. So, as we move around, we leak our